Daily reading queue

Topology-Enhanced Alignment for Large Language Models: Trajectory Topology Loss and Topological Preference Optimization

Yurui Pan, Ke Xu, Bo Peng · 2026-05-12

The authors propose using topological methods (specifically 0-dimensional persistent homology, which identifies connected components and their lifetimes across scales) to improve LLM alignment. They introduce two techniques: Trajectory Topology Loss (TTL) for supervised fine-tuning, which regularizes the model's update direction to follow "prompt-answer bridges" extracted from the geometry of embeddings, and Topological Preference Optimization (TPO) for DPO, which aligns the improvement direction from rejected to chosen responses with topic-specific semantic vectors. Testing on Qwen2.5-7B-Instruct shows consistent improvements over baselines on preference metrics and LLM-judge evaluations. **Main takeaways:** - Views generation as tracing a trajectory through hidden representation space, not just a sequence of token probabilities - TTL treats prompt and answer embeddings as a point cloud, uses persistent homology to find topological structure, and aligns model updates with that structure - TPO constructs semantic preference vectors for topics and guides the DPO update direction in intermediate layers to align with those vectors - Outperforms non-topological baselines (per-example, nearest-neighbor, random regularizers) on automatic metrics and GPT-judge evaluation - Maintains or improves toxicity scores while improving preference alignment

arxiv cs.CL (NLP)

arXiv:2605.07172v1 Announce Type: new Abstract: Alignment of large language models (LLMs) via SFT and RLHF/DPO typically ignores the global geometry of the representation space, relying instead on local token likelihoods or scalar scores. We view generation as tracing a semantic trajectory in hidden space and propose a topology-enhanced alignment framework that regularizes these trajectories using 0-dimensional persistent homology. First, for SFT, we introduce Trajectory Topology Loss (TTL). Treating prompt and gold-answer embeddings as a mixed point cloud, we use a 0D persistent homology algorithm to extract "prompt-answer bridges." TTL aligns the model's actual update direction with these topological bridges rather than arbitrary directions. Second, for DPO, we propose Topological Preference Optimization (TPO). TPO constructs topic-specific semantic preference vectors and aligns the improvement direction between rejected and chosen responses with these vectors in an intermediate hidden layer. We also introduce a dynamic weighting scheme to balance DPO and TPO losses. Evaluating on Qwen2.5-7B-Instruct using UltraChat and Anthropic HH-RLHF, our topology-enhanced objectives consistently outperform strong non-topological baselines (e.g., per-example, nearest-neighbor, random regularizers) on automatic preference metrics and LLM-judge evaluations, while maintaining or improving toxicity. Results show persistent homology and trajectory geometry offer a promising direction for controllable alignment.

Potential threat/caveat for clean result "Stretching turn count, completion length, or system-prompt length at train time fails to amplify marker uptake; the longest system prompt instead leaks across bystander personas (LOW confidence)": this item discusses evaluation.

A Reproducible Multi-Architecture Baseline for Token-Level Chinese Metaphor Identification under the MIPVU Framework

Yufeng Wu · 2026-05-12

This paper establishes a reproducible baseline for identifying metaphor-related words in Chinese text at the token level, following the MIPVU linguistic framework. The author compares three approaches on the PSU Chinese Metaphor Corpus: fine-tuned Chinese RoBERTa, MelBERT (adapted to Chinese with a new basic-meaning resource from a Chinese dictionary covering 71.51% of corpus vocabulary), and instruction-tuned Qwen3.5-9B. MelBERT achieves the best F1 (0.7281), marginally beating plain RoBERTa (0.7142), while Qwen lags by ~11 F1 points (0.6157) due to recall problems from discrete output formatting. **Main takeaways:** - MelBERT's basic-meaning channel (which checks if a word's usage matches its dictionary definition) works for Chinese metaphor detection, unlike the "SPV" channel (checking if novel meanings emerge) - The generative Qwen model struggles with recall because it must commit to discrete labels rather than producing continuous scores - Several Qwen task formulations failed due to output format design issues, not model capacity limits - The author releases all splits, outputs, the Chinese basic-meaning resource (from Modern Chinese Dictionary 7th edition), and training scripts - Conventional metaphor dominates in Chinese, consistent with why the SPV (novel-meaning) channel doesn't help much

arxiv cs.CL (NLP)

arXiv:2605.07170v1 Announce Type: new Abstract: Metaphor is pervasive in everyday language, yet token-level computational identification of metaphor-related words in Chinese under the MIPVU framework remains under-explored relative to English. This paper presents a reproducible multi-architecture baseline for token-level metaphor identification on the PSU Chinese Metaphor Corpus (PSU CMC), the only widely available MIPVU-annotated Chinese corpus. We systematically compare three model families: (i) encoder fine-tuning with Chinese RoBERTa-wwm-ext-large; (ii) MelBERT adapted to Chinese using a newly constructed basic-meaning resource derived from the Modern Chinese Dictionary, 7th edition (MCD7), comprising 74,823 entries with 71.51% PSU CMC vocabulary coverage; and (iii) Qwen3.5-9B fine-tuned with QLoRA as an instruction-tuned generative baseline. Across five fixed seeds, MelBERT MIP-only achieves the strongest performance at 0.7281 +/- 0.0050 test positive F1, marginally above MelBERT Full (0.7270 +/- 0.0069) and clearly above plain RoBERTa (0.7142 +/- 0.0121). The Qwen QLoRA generative configuration trails encoder baselines by approximately 11 F1 points (0.6157 +/- 0.0113). Three findings merit attention: (1) the SPV channel of MelBERT does not contribute reliable positive signal in Chinese, consistent with the dominance of conventional metaphor; (2) the Qwen-encoder gap is concentrated in recall, reflecting the discrete-commitment limitation of generative output; (3) several Qwen task formulations fail due to format design rather than model capacity. We release all split manifests, per-seed outputs, the MCD7 basic-meaning embedding pipeline, and training scripts to serve as a common reference for future Chinese metaphor identification research.

Potential threat/caveat for clean result "Fine-tuning one persona on a two-marker chunk and another on the start marker plants the end marker at every donor answer's end, not chained to the start (LOW confidence)": this item discusses limitation.

Rethinking Experience Utilization in Self-Evolving Language Model Agents

Weixiang Zhao, Yingshuo Wang, Yichen Zhang, Yanyan Zhao, Yu Zhang, Yang Wu, Dandan Tu, Bing Qin, Ting Liu · 2026-05-12

The authors study when agents should use accumulated experience during decision-making, not just how experience should be stored or updated. They introduce ExpWeaver, which exposes experience as an optional resource during reasoning so agents can invoke it only when needed, rather than injecting it at every step or only at initialization. Across four agent frameworks, seven LLM backbones, and three environments, ExpWeaver consistently outperforms fixed usage strategies. Analysis shows agents learn to invoke experience selectively at beneficial decision points and under higher reasoning uncertainty. **Main takeaways:** - Most self-evolving agents use rigid experience-injection strategies (always at start, or always at every step), ignoring whether experience is actually needed - ExpWeaver makes experience optional during reasoning, letting the agent decide when to retrieve and use it - Consistently achieves best performance across diverse frameworks, model sizes, and task types - RL training amplifies the selective-invocation behavior - Usage-pattern and entropy analyses show agents invoke experience when facing higher decision uncertainty and at beneficial choice points - Suggests experience utilization (when to use stored knowledge) is as important as experience construction (what to store)

arxiv cs.CL (NLP)

arXiv:2605.07164v1 Announce Type: new Abstract: Self-evolving agents improve by accumulating and reusing experience from past interactions. Existing work has largely focused on how experience is constructed, represented, and updated, while paying less attention to how experience should be used during runtime decision-making. As a result, most agents rely on rigid usage strategies, either injecting experience once at initialization or at every step, without considering whether it is needed for the current decision. This paper studies experience utilization as a critical design dimension of self-evolving agents. We ask whether agents benefit from interweaving experience use with decision-making, so that experience is invoked only when additional guidance is needed. To examine this question, we introduce {ExpWeaver}, a lightweight instantiation that leaves experience construction unchanged and modifies only runtime utilization by exposing experience as an optional resource during reasoning. Across four representative frameworks, seven LLM backbones, and three types of environments, ExpWeaver consistently achieves the best performance among different utilization strategies. Reinforcement learning experiments further show that this behavior can be amplified through training. Usage-pattern, causal ablation, and entropy-based analyses reveal that ExpWeaver enables agents to invoke experience selectively, at beneficial decision points, and under higher reasoning uncertainty. Overall, our findings call for a shift from merely studying \emph{what} experience to store toward understanding \emph{how} and \emph{when} experience should enter decision-making.

CLIPer: Tailoring Diverse User Preference via Classifier-Guided Inference-Time Personalization

Jinyan Su, Jinpeng Zhou, Claire Cardie, Wen Sun · 2026-05-12

CLIPer uses a lightweight classifier to steer LLM generation at inference time toward different user preferences (helpfulness, conciseness, humor, etc.) without fine-tuning a separate model for every preference combination. The classifier guides generation dynamically, adding negligible computational cost while enabling controllable personalization across single and multi-dimensional preferences. Empirical results show the approach scales well and delivers effective personalized generation without extensive training. **Main takeaways:** - Eliminates the need to fine-tune separate models for each preference combination (helpfulness+concise, helpful+humorous, etc.) - Uses a classifier model to dynamically steer generation toward desired preferences at inference time - Works across single preferences (just conciseness) and multi-dimensional combinations (concise + helpful) - Adds negligible computational overhead compared to fine-tuning multiple models - Enables more nuanced control over generation style without retraining

arxiv cs.CL (NLP)

arXiv:2605.07162v1 Announce Type: new Abstract: Personalized LLMs can significantly enhance user experiences by tailoring responses to preferences such as helpfulness, conciseness, and humor. However, fine-tuning models to address all possible combinations of user preferences is computationally expensive and impractical. In this paper, we introduce \textbf{CLIPer}(\textbf{Cl}assifier-guided \textbf{I}nference-time \textbf{Per}sonalization), a lightweight personalization approach that leverages a classifier model to steer LLM generation dynamically to different user preferences at inference time. Our method eliminates the need for extensive fine-tuning, inducing negligible additional computational overhead while enabling more controllable and nuanced personalization across single and multi-dimensional preferences. Comprehensive empirical analyses demonstrate the scalability and effectiveness of our approach in delivering personalized language generation.

Beyond Reasoning: Reinforcement Learning Unlocks Parametric Knowledge in LLMs

Wanli Yang, Hongyu Zang, Junwei Zhang, Wenjie Shi, Du Su, Jingang Wang, Xueqi Cheng, Fei Sun · 2026-05-12

The authors use reinforcement learning (RL) on binary correctness rewards to improve a language model's ability to recall factual knowledge directly, without any chain-of-thought reasoning. Testing on closed-book QA, they get roughly 27% relative improvement across models. Mechanistically, RL doesn't teach new facts—it moves correct answers that already exist somewhere in the model's outputs (often in the low-probability tail) up into the top greedy predictions. The hardest training examples (where the right answer appeared in fewer than 1 in 128 pre-RL samples) drive 83% of the improvement, because even rare correct rollouts get reinforced. **Main takeaways:** - RL on simple correctness rewards improves factual recall by ~27% without reasoning chains or memorization of training data - The mechanism is probability redistribution: moving existing correct answers from rare samples to greedy top-1 outputs - The hardest examples (18% of training data) contribute 83% of the gain because their occasional correct rollouts get amplified - RL acts as a tool for "unlocking" latent knowledge rather than installing new facts

arxiv cs.CL (NLP)

arXiv:2605.07153v1 Announce Type: new Abstract: Reinforcement learning (RL) has achieved remarkable success in LLM reasoning, but whether it can also improve direct recall of parametric knowledge remains an open question. We study this question in a controlled zero-shot, one-hop, closed-book QA setting with no chain-of-thought, training only on binary correctness rewards and applying fact-level train-test deduplication to ensure gains reflect improved recall rather than reasoning or memorization. Across three model families and multiple factual QA benchmarks, RL yields ~27% average relative gains, surpassing both training- and inference-time baselines alike. Mechanistically, RL primarily redistributes probability mass over existing knowledge rather than acquiring new facts, moving correct answers from the low-probability tail into reliable greedy generations. Our data-attribution study reveals that the hardest examples are the most informative: those whose answers never appear in 128 pre-RL samples (only ~18% of training data) drive ~83% of the gain, since rare correct rollouts still emerge during training and get reinforced. Together, these findings broaden the role of RL beyond reasoning, repositioning it as a tool for unlocking rather than acquiring latent parametric knowledge.

Potential threat/caveat for clean result "Training a `[ZLT]` persona-marker into Qwen-2.5-7B doesn't increase system-prompt attention at the marker timestep — base Qwen on identical tokens attends the same way (LOW confidence)": this item discusses benchmark.

Structural Rationale Distillation via Reasoning Space Compression

Jialin Yang, Jiankun Wang, Jiajun Wu, Henry Leung, Jiayu Zhou, Steve Drew · 2026-05-12

When distilling reasoning from a large teacher model into a smaller student, the teacher's step-by-step explanations for similar problems often vary wildly, creating noisy supervision. The authors propose D-RPC, which maintains a compact bank of reusable high-level reasoning "paths" (templates) and forces the teacher to follow the most relevant one for each question, producing more consistent rationales. A PAC-Bayes analysis shows an optimal intermediate bank size: too small misses coverage, too large adds noise. Across five reasoning benchmarks, D-RPC beats standard chain-of-thought distillation and other baselines while using fewer tokens. **Main takeaways:** - Teacher models generate inconsistent reasoning paths for similar problems, which hurts student learning during distillation - Constraining the teacher to follow retrieved high-level reasoning templates improves student performance - There's a Goldilocks bank size: too few templates miss problem types, too many add supervision noise - D-RPC outperforms chain-of-thought distillation and other methods on math and commonsense benchmarks with lower token cost

arxiv cs.CL (NLP)

arXiv:2605.07139v1 Announce Type: new Abstract: When distilling reasoning from large language models (LLMs) into smaller ones, teacher rationales for similar problems often vary wildly in structure and strategy. Like a chef who makes the same dish differently each time, this inconsistency burdens the student with noisy supervision that is hard to internalize. We propose Distillation through Reasoning Path Compression (D-RPC), which constrains the teacher to follow a compact, dynamically maintained bank of reusable high-level reasoning paths. For each training question, D-RPC retrieves the most relevant path and conditions the teacher to follow it, producing rationales that are consistent across similar problems yet diverse enough to cover different problem types. A PAC-Bayes analysis formalizes the resulting trade-off between bank size and coverage: smaller banks reduce supervision entropy but risk coverage gaps, and the generalization bound identifies an optimal intermediate size confirmed by our ablations. Across five math and commonsense reasoning benchmarks with two student models, D-RPC consistently outperforms chain-of-thought distillation, freeform rationale generation, direct distillation, and structured-supervision baselines, while using fewer tokens than template-heavy alternatives.

Potential threat/caveat for clean result "Fine-tuning one persona on a two-marker chunk and another on the start marker plants the end marker at every donor answer's end, not chained to the start (LOW confidence)": this item discusses benchmark.

Region4Web: Rethinking Observation Space Granularity for Web Agents

Donguk Kwon, Dongha Lee · 2026-05-12

Web agents typically observe web pages at the granularity of individual elements (buttons, links, etc.), which forces them to infer functional organization at every step. The authors propose Region4Web, which groups elements into functional regions (parts of the page serving distinct purposes) and presents these as a compact "digest" that persists across reasoning steps. On WebArena, this region-level observation reduces input length and improves task success across different LLMs and agent architectures, regardless of model size. **Main takeaways:** - Current web agents observe pages at element-level granularity, leaving functional structure implicit - Grouping elements into functional regions (e.g., navigation bar, search box, content area) gives agents a better basis for understanding page state - PageDigest delivers this as a persistent per-page summary rather than re-processing elements every step - Shorter observations and higher success rates across LLM backbones on WebArena benchmark

arxiv cs.CL (NLP)

arXiv:2605.07134v1 Announce Type: new Abstract: Web agents perceive web pages through an observation space, yet its granularity has remained an underexamined design choice. Existing work treats observation at the same element-level granularity as the action space, leaving the page's functional organization implicit and forcing the agent to infer it from element-level signals at every step. We argue observation should instead operate at the granularity of functional regions, parts of the page that each serve a distinct purpose. We propose Region4Web, a framework that reorganizes the AXTree into functional regions through hierarchical decomposition and semantic abstraction, exposing the page's functional organization as the basis for page state understanding. Moreover, we propose PageDigest, a web-specific inference pipeline that delivers this region-level observation to the actor agent as a compact per-page digest that persists across steps. On the WebArena benchmark, PageDigest substantially reduces observation length while improving overall task success rate across diverse backbone large language models (LLMs) and established agent methods, regardless of backbone capacity. These results show that operating at the granularity of functional regions delivers a more compact and informative basis for the actor agent than element-level processing alone.

Potential threat/caveat for clean result "Fine-tuning one persona on a two-marker chunk and another on the start marker plants the end marker at every donor answer's end, not chained to the start (LOW confidence)": this item discusses benchmark.

Beyond LoRA vs. Full Fine-Tuning: Gradient-Guided Optimizer Routing for LLM Adaptation

Haozhan Tang, Xiuqi Zhu, Xinyin Zhang, Boxun Li, Virginia Smith, Kevin Kuo · 2026-05-12

Full fine-tuning (FFT) gives models the plasticity to learn high-entropy new knowledge, while LoRA often matches or beats FFT because many tasks only need low-rank updates and benefit from LoRA's regularization. The authors propose MoLF, which dynamically routes gradient updates between FFT and LoRA experts at the optimizer level during training, letting the model use whichever is better for each update. Across SQL, medical QA, and counterfactual knowledge tasks, MoLF matches or beats the better of FFT or LoRA by up to 1.5%. A memory-efficient variant (MoLF-Efficient) freezes base weights and routes between two LoRA experts of different rank. **Main takeaways:** - Full fine-tuning provides plasticity for complex updates; LoRA provides regularization and works when updates are low-rank - Statically choosing one method is suboptimal—different tasks and training steps benefit from different regimes - MoLF routes updates between FFT and LoRA experts dynamically at the optimizer level, getting the best of both - MoLF-Efficient (LoRA-only routing) outperforms prior adaptive LoRA methods by up to 20% on factual tasks

arxiv cs.CL (NLP)

arXiv:2605.07111v1 Announce Type: new Abstract: Recent literature on fine-tuning Large Language Models highlights a fundamental debate. While Full Fine-Tuning (FFT) provides the representational plasticity required for high-entropy knowledge injection, Low-Rank Adaptation (LoRA) can match or surpass FFT performance because many tasks only require updates in a low-rank space and benefit from LoRA's additional regularization. Through empirical evaluation across diverse tasks (SQL, Medical QA, and Counterfactual Knowledge) and varying language models (Gemma-3-1B, Qwen2.5-1.5B, and Qwen2.5-3B), we verify both trends and demonstrate that relying solely on either static architecture is structurally limited. To address this challenge, we propose a Mixture of LoRA and Full (MoLF) Fine-Tuning, a unified framework that enables continuous navigation between both training regimes. MoLF dynamically routes updates between FFT and LoRA at the optimizer level to ensure that exact gradient signals are available to both experts throughout training, yielding stable training dynamics. For memory-constrained environments, we also introduce MoLF-Efficient, which freezes base weights and only routes updates among a pair of LoRA experts of potentially varying rank. Our evaluations show that MoLF either improves on or stays within $1.5\%$ of the better of FFT and LoRA across all settings, while MoLF-Efficient outperforms prior adaptive LoRA approaches by up to $20\%$ on Fact and $9\%$ on Med and SQL.

Potential threat/caveat for clean result "Fine-tuning one persona on a two-marker chunk and another on the start marker plants the end marker at every donor answer's end, not chained to the start (LOW confidence)": this item discusses evaluation.

Securing Computer-Use Agents: A Unified Architecture-Lifecycle Framework for Deployment-Grounded Reliability

Zejian Chen, Zhanyuan Liu, Chaozhuo Li, Mengxiang Han, Songyang Liu, Litian Zhang, Feng Gao, Yiming Hei, Xi Zhang · 2026-05-12

Computer-use agents (CUAs) that control browsers, terminals, and applications face reliability challenges beyond task success: perception errors, planning drift, permission scope, and runtime oversight all affect whether actions stay aligned with user intent. This paper develops a framework organized by architectural layers (Perception → Decision → Execution) and lifecycle stages (Creation → Deployment → Operation → Maintenance) to analyze where failures are introduced versus where they become visible, and maps intervention surfaces for control and assurance. **Main takeaways:** - Agent reliability isn't just task success—it includes perception accuracy, planning stability, memory use, tool safety, and permission boundaries - The framework separates architectural layers (how agents transform observations into actions) from lifecycle stages (when priors are learned, tools bound, and drift occurs) - Failures often become visible in different stages than where their root causes are introduced - Identifies open challenges: controllable grounding, long-horizon constraint preservation, safe authority binding, runtime defense

arxiv cs.CL (NLP)

arXiv:2605.07110v1 Announce Type: new Abstract: Computer-use agents(CUAs)are moving frombounded benchmarks toward real software environments, wherethey operate browsers, desktops, mobile applications, flesystems,terminals, and tool backends. In such settings, reliability isno longer captured by task success alone: perception errors,planning drift, memory use, tool mediation, permission scope,and runtime oversight jointly determine whether agent actionsremain aligned with user intent, Existing surveys organize theCUA landscape by methods, platforms, benchmarks, or securitythreats, but less explicitly connect capability formation, author-ity exposure, failure manifestation, and control placement. Toaddress this gap, the article develops an architecture-lifecycleframework for deployment-grounded reliability in CUAs. Thearchitectural view analyzes Perception, Decision, and Executionas coupled layers that transform software observations intoauthority-bearing actions, The lifecycle view examines Creation.Deployment, Operation, and Maintenance as stages in which priorsare learned, tools and permissions are bound, runtime trajecto.ries are stressed, and assurance must be preserved under drift.Using this lens, the analysis synthesizes representative systems,benchmarks, and security/privacy studies; distinguishes wherefailures become visible from where their enabling conditions areintroduced, and maps recurring intervention surfaces for controloversight, and assurance. OpenClaw is used only as a public moti.vating example of an open deployment pattern, not as a verifedinternal case study. The conclusion highlights open challengesin controllable grounding, long-horizon constraint preservation,safe authority binding, mixed-trust runtime defense, privacy-preserving memory,and continual assurance.

Potential threat/caveat for clean result "Fine-tuning one persona on a two-marker chunk and another on the start marker plants the end marker at every donor answer's end, not chained to the start (LOW confidence)": this item discusses failure, failures, benchmark.

Retrieve, Integrate, and Synthesize: Spatial-Semantic Grounded Latent Visual Reasoning

Jin Cui, Xinyue Long, Xunyong Zhang, Yadong Zhang, Chuanchang Su, Jingye Gan, Boran Zhao, Pengju Ren · 2026-05-12

Multimodal models that compress visual reasoning into text lose fine-grained information. Recent "latent reasoning" methods try to reason in continuous hidden states instead, but the authors find these latent trajectories drift from the model's pretrained circuits, collapse into generic patterns, and get bypassed during answer generation. They propose RIS, which grounds latent reasoning tokens to spatial bounding boxes and semantic region descriptions, enforces their causal role through a progressive attention bottleneck, and uses short language tokens to bridge back to vocabulary-aligned decoding. RIS improves over baselines on vision-reasoning benchmarks and produces more interpretable latent trajectories. **Main takeaways:** - Latent visual reasoning (reasoning in hidden states rather than text) often fails because trajectories drift from pretrained circuits and get ignored during decoding - RIS anchors latent tokens to spatial (bounding boxes) and semantic (region descriptions) evidence to keep them grounded - A progressive attention bottleneck forces the model to actually use latent tokens rather than bypass them - Short language "transition tokens" help bridge continuous latent states back to vocabulary-aligned text generation

arxiv cs.CL (NLP)

arXiv:2605.07106v1 Announce Type: new Abstract: Multimodal Large Language Models (MLLMs) have made remarkable progress on vision-language reasoning, yet most methods still compress visual evidence into discrete textual thoughts, creating an information bottleneck for fine-grained perception. Recent latent visual reasoning methods attempt to reason in continuous hidden states, but we find that they suffer from insufficient manifold compatibility: latent trajectories drift away from pretrained reasoning circuits, collapse into instance-agnostic patterns, and are often bypassed during answer generation. To address these issues, we propose RIS (Retrieve, Integrate, and Synthesize), a spatial-semantic grounded framework that develops latent reasoning as a compatible extension of pretrained MLLM computation. We first construct a step-wise grounded reasoning dataset with bounding boxes and region-specific semantic descriptions. Built on this supervision, RIS anchors latent tokens to both spatial and semantic evidence, enforces their causal role through a progressive attention bottleneck, and introduces short language transition tokens to bridge synthesized latent states back to vocabulary-aligned decoding. Experiments on V*, HRBench4K, HRBench8K, MMVP, and BLINK show consistent improvements over closed/open-source and latent reasoning baselines. Further analyses demonstrate that RIS learns diverse, interpretable, and progressively integrated latent trajectories, offering a practical path toward faithful internal visual reasoning in MLLMs.

SAGE: Hierarchical LLM-Based Literary Evaluation through Ontology-Grounded Interpretive Dimensions

Tianyu Wang, Nianjun Zhou · 2026-05-12

Evaluating literary quality requires assessing interpretive dimensions (cultural representation, emotional depth, philosophical sophistication) that are hard to measure computationally. The authors introduce SAGE, which uses LLMs to evaluate stories across three hierarchical layers (cultural, emotional-psychological, existential-philosophical) with iterative reflection and independent validation. Testing on 100 stories (canonical, pulp fiction, LLM-generated), they achieve 98.8% score convergence and >94% inter-rater agreement. Canonical works consistently outperform pulp and LLM-generated stories on cultural and philosophical dimensions (effect sizes >2.4), but emotional representation shows smaller gaps (d=1.68), suggesting affective patterns are more learnable from training data. **Main takeaways:** - LLM-based evaluation can achieve measurement-grade reliability on complex interpretive dimensions with structured prompts and reflection - Canonical literature > pulp > LLM-generated on cultural and philosophical depth, with large effect sizes (d>2.4) - Emotional representation shows smaller gaps (d=1.68), suggesting current LLMs learn affective patterns better than critical stance - Three quality dimensions (cultural, emotional, philosophical) are empirically distinguishable and moderately correlated (r=0.65-0.68)

arxiv cs.CL (NLP)

arXiv:2605.07102v1 Announce Type: new Abstract: Evaluating literary quality requires assessing interpretive dimensions such as cultural representation, emotional depth, and philosophical sophistication that resist straightforward computational measurement. We introduce SAGE, a hierarchical evaluation framework that decomposes literary quality into ontology-grounded interpretive dimensions assessed through structured large language model evaluation with multi-round iterative reflection and independent validation. We validate the framework on 100 short stories (50 canonical works, 30 pulp fiction, 20 LLM-generated narratives) across three analytical layers (cultural, emotional-psychological, existential-philosophical) using dual-mode assessment. Across 600 evaluations, the framework achieves 98.8% score convergence and greater than 94% inter-rater agreement, with near-perfect mode invariance between content-based and metadata-based evaluation. Statistical analysis reveals a consistent genre hierarchy (Canonical > Pulp > LLM, all p2.4), while emotional representation shows smaller gaps (d=1.68), suggesting that affective patterns are more learnable from training data than critical stance or philosophical depth. Cross-layer correlations (r=0.649-0.683) confirm the three dimensions capture empirically distinguishable quality facets. These findings demonstrate that theory-driven LLM evaluation can achieve measurement-grade reliability and support systematic identification of where current generative models fall short of human literary production, with direct implications for scalable automated evaluation of open-ended text generation.

Potential threat/caveat for clean result "Fine-tuning one persona on a two-marker chunk and another on the start marker plants the end marker at every donor answer's end, not chained to the start (LOW confidence)": this item discusses evaluation.

The Translation Tax Is Not a Scalar: A Counterfactual Audit of English-Source Cue Inheritance in Chinese Multilingual Benchmarks

Zezheng Lin, Fengming Liu, Handi Li · 2026-05-12

Translated benchmarks are assumed to inflate scores by preserving English-source cues (the "Translation Tax"), but the authors show this isn't a simple scalar effect when auditing English-to-Chinese benchmarks. Three different measurement approaches disagree: back-translation gaps are small, cue-score calibration doesn't predict item-level gains, and native-control comparisons show model-family effects rather than uniform inflation. An LLM-naturalization stress test (rewriting Chinese surface form while keeping content fixed) reveals item-dependent validity risks rather than a single translation tax—high-residue items benefit from naturalization, low-residue items don't. **Main takeaways:** - The "Translation Tax" (score inflation from English cues in translated benchmarks) isn't a uniform scalar effect - Three proxy estimators give inconsistent results: back-translation, cue-score calibration, and native-control comparisons - LLM naturalization (rewriting surface form while preserving content) shows item-dependent effects: some items benefit, others don't - Translation validity risk depends on estimator choice, item properties, and model family, not a single benchmark-wide factor

arxiv cs.CL (NLP)

arXiv:2605.07093v1 Announce Type: new Abstract: The Translation Tax is often treated as a scalar: translated benchmarks are assumed to inflate scores by preserving English-source cues. We audit this claim in an English-to-Chinese setting. Three proxy estimators disagree: back-translation gaps are small and parser-fragile; cue-score calibration does not predict item-level gains; and a six-model native-control comparison shows model-family rather than uniform benchmark effects. We add a same-item LLM-naturalization stress test that holds answer, options, and content fixed while rewriting Chinese surface form. After correcting a prompt-construction bug, this contrast no longer supports a model-family interaction, but it preserves a residue dose-response: high-residue items benefit while low-residue items do not. The result is not a single Translation Tax, but a set of estimator- and item-dependent validity risks. We release per-cell evidence, the naturalization protocol, human QC, and a reporting checklist for translated multilingual benchmark papers.

Potential threat/caveat for clean result "Stretching turn count, completion length, or system-prompt length at train time fails to amplify marker uptake; the longest system prompt instead leaks across bystander personas (LOW confidence)": this item discusses benchmark.

Beyond Single Ground Truth: Reference Monism as Epistemic Injustice in ASR Evaluation

Anna Seo Gyeong Choi, Maria Teleki, James Caverlee, Miguel del Rio, Corey Miller, Hoon Choi · 2026-05-12

The authors argue that automatic speech recognition (ASR) systems commit "epistemic injustice" when they enforce a single transcription standard as ground truth. Different transcription conventions (verbatim vs. cleaned-up) produce different "correct" transcripts for the same speech, and speakers with aphasia—whose disfluencies carry clinical meaning—are systematically penalized when evaluated against "clean" references that treat those features as errors. The paper proposes WER-Range: reporting performance across multiple legitimate transcription conventions instead of assuming one right answer. **Main takeaways:** - Word Error Rate (WER) varies depending on which transcription convention you pick as ground truth—not because the system changed, but because "ground truth" itself is a choice. - Speakers with aphasia are disadvantaged when their meaningful disfluencies are treated as errors to be removed from the reference transcript. - The authors introduce Epistemic Injustice Distance (EID) to quantify the harm of enforcing a single standard. - WER-Range proposes reporting ASR performance across multiple legitimate conventions rather than picking one. - The problem isn't just differential performance—it's that the evaluation infrastructure lacks the conceptual tools to recognize some speech as legitimate in the first place.

arxiv cs.CL (NLP)

arXiv:2605.07084v1 Announce Type: new Abstract: Automatic speech recognition (ASR) evaluation compares system output to ground truth transcripts, with Word Error Rate (WER) quantifying the distance between them. But ground truth transcripts are not discovered - they are produced by human annotators following conventions that encode normative assumptions about which speech features matter. Different conventions (verbatim, non-verbatim, legal) produce different transcripts of identical speech and judge the same ASR output differently. This paper argues that reference monism - enforcing a single transcription convention as ground truth - commits epistemic injustice. Speakers with aphasia, whose speech includes clinically meaningful disfluencies, are systematically disadvantaged when evaluated against "clean" references that treat those disfluencies as errors. The harm is not merely differential performance, but that evaluative infrastructure lacks interpretive resources to recognize their contributions as legitimate. We develop a philosophical framework introducing the hermeneutical gap, formalize Epistemic Injustice Distance (EID) to measure reference monism's cost, and demonstrate empirically using AphasiaBank that WER varies depending on which convention defines ground truth. We propose WER-Range: reporting performance across legitimate conventions rather than assuming a single correct answer.

Potential threat/caveat for clean result "Fine-tuning one persona on a two-marker chunk and another on the start marker plants the end marker at every donor answer's end, not chained to the start (LOW confidence)": this item discusses evaluation.

Self-Consolidating Language Models: Continual Knowledge Incorporation from Context

Zekun Wang, Anant Gupta, Zihan Dong, Christopher J. MacLellan · 2026-05-12

The authors propose Self-Consolidating Language Models (SCoL), a framework that lets an LLM decide which of its own Transformer layers to update when incorporating new context. Instead of simply putting information in the prompt or fine-tuning everything, the model learns to generate "update instructions" specifying sparse layer selections, trained with meta-reinforcement learning so it can adapt as its own weights change. On QA tasks and long-context benchmarks, SCoL beats prompting, summarization, and sequential fine-tuning by learning to route updates toward high-Fisher-information layers (the parts of the model most sensitive to loss) while limiting interference with previously consolidated knowledge. **Main takeaways:** - SCoL lets the model choose which layers to update when incorporating new context, rather than updating all weights or none. - Trained via meta-RL over an evolving model state—the model that picks future updates is itself being changed by past updates. - Outperforms prompting, summarization, batch test-time training, and sequential fine-tuning on both SQuAD knowledge incorporation and LongBench v2 long-context tasks. - Learned update patterns are sparse and align with layers of high Fisher information (regions where small weight changes have big effects on loss). - Transfers from shorter training streams to longer evaluation streams, suggesting the method scales to streaming contexts.

arxiv cs.CL (NLP)

arXiv:2605.07076v1 Announce Type: new Abstract: Large language models (LLMs) increasingly receive information as streams of passages, conversations, and long-context workflows. While longer context windows expose more evidence, they do not ensure that useful information is preserved and reused. We study continual context consolidation: writing current context into model weights while limiting interference with previously consolidated information. We propose \textbf{S}elf-\textbf{Co}nsolidating \textbf{L}anguage Models (SCoL), a post-training framework in which, given current context, an LLM learns to generate textual update instructions specifying which of its own Transformer layers should be updated. Because committed updates change the model that later generates future selections, we train SCoL with meta-reinforcement learning over an evolving model state. We instantiate SCoL with supervised QA rewards on SQuAD knowledge incorporation and intrinsic likelihood-based rewards for LongBench v2 long-context consolidation. Across both settings, SCoL improves acquisition and retention over prompting, summarization, batch test-time training, and sequential finetuning baselines. Analysis of learned selection patterns shows that SCoL encourages the LLM to generate sparse update locations that align with layers of high Fisher information, suggesting that the model learns to route plasticity toward loss-sensitive regions while limiting interference. Moreover, SCoL transfers from shorter meta-training streams to longer LongBench v2 streams at evaluation, suggesting that our framework supports scalable streaming consolidation.

Potential threat/caveat for clean result "Fine-tuning one persona on a two-marker chunk and another on the start marker plants the end marker at every donor answer's end, not chained to the start (LOW confidence)": this item discusses evaluation.

WiCER: Wiki-memory Compile, Evaluate, Refine Iterative Knowledge Compilation for LLM Wiki Systems

Juan M. Huerta · 2026-05-12

The authors tackle the "LLM Wiki" pattern—compiling domain knowledge into a persistent artifact served via KV cache for sub-second, zero-retrieval-failure access. The core problem is the "compilation gap": blindly summarizing documents into a wiki catastrophically drops facts (53–60% failure rate), while serving the full raw context works well but doesn't scale due to attention dilution. They propose WiCER (Wiki-memory Compile, Evaluate, Refine), an iterative algorithm inspired by counterexample-guided refinement that evaluates compiled wikis against diagnostic questions, identifies dropped facts, and forces their preservation in the next compilation round. One or two iterations recover 80% of lost quality and reduce catastrophic failures by 55%. **Main takeaways:** - Full-context KV cache inference beats RAG on curated knowledge (4.38 vs. 4.08 out of 5, 7× faster) but degrades at scale due to attention dilution. - Blind compilation into a wiki fails catastrophically: drops quality from ~3.5 to ~2.2 and causes 53–60% failure rates. - WiCER iteratively diagnoses dropped facts and forces the next compilation to preserve them, recovering 80% of lost quality in 1–2 rounds. - Targeted diagnosis (identifying specific missing facts) drives the gains (+0.95); generic "pinning" strategies only help marginally (+0.16). - All code and benchmarks are released across 17 RepLiQA domains (6,800 questions).

arxiv cs.CL (NLP)

arXiv:2605.07068v1 Announce Type: new Abstract: The LLM Wiki pattern, to compile and provide domain knowledge into a persistent artifact and serve it to LLMs via KV cache inference, promises context access at sub-second latency with zero retrieval failure. Realizing this requires solving the compilation gap: LLM compilation distilling raw documents into a wiki without catastrophically discarding critical facts. We characterize this gap across 17 RepLiQA domains (6,800 questions): we observe that full context KV cache inference outperforms RAG on curated knowledge (4.38 vs. 4.08 out of 5, 7.3 faster TTFT) but degrades below RAG at scale due to attention dilution, and blind compilation fails entirely (2.14 to 2.32 vs. 3.46, 53 to 60% catastrophic failure rate). To address the compilation gap, we propose WiCER (Wiki-memory Compile, Evaluate, Refine), an iterative algorithm inspired by counterexample-guided abstraction refinement (CEGAR) that closes this gap. WiCER evaluates compiled wikis against diagnostic probes, identifies dropped facts, and forces their preservation in subsequent compilations. One to two iterations recover 80% of lost quality (mean 3.24 vs. 3.47 for raw full-context across the 15 topics with baselines), reducing catastrophic failures by 55% relative. An ablation across all 17 topics confirms that targeted diagnosis (+0.95), not generic pinning (+0.16), drives the gains. All code and benchmarks are released for reproducible research.

Potential threat/caveat for clean result "Fine-tuning one persona on a two-marker chunk and another on the start marker plants the end marker at every donor answer's end, not chained to the start (LOW confidence)": this item discusses failure, failures, counterexample, benchmark.

MedExAgent: Training LLM Agents to Ask, Examine, and Diagnose in Noisy Clinical Environments

Yicheng Gao, Xiaolin Zhou, Yahan Li, Yue Zhao, Ruishan Liu · 2026-05-12

The authors formalize clinical diagnosis as a Partially Observable Markov Decision Process (POMDP) where the agent can ask questions, order exams (as tool calls), or issue a diagnosis—all under systematic noise (seven patient noise types, three exam noise types). Existing medical LLM benchmarks simplify this to single-turn QA or noise-free conversations, missing the interactive, uncertain nature of real diagnosis. They train MedExAgent via supervised fine-tuning on synthetic Calgary-Cambridge-style clinical interviews, then apply DAPO (a policy optimization method) to maximize a composite reward: diagnostic accuracy, tool-call quality, and exam cost (financial + patient discomfort). MedExAgent matches larger models' diagnostic performance while maintaining cost-efficient exam strategies. **Main takeaways:** - Real clinical diagnosis involves questioning, exam ordering, and diagnosis under noisy, incomplete information—existing benchmarks ignore this interactivity. - MedExAgent formalizes diagnosis as a POMDP with three action types and a systematic noise model (e.g., vague patient answers, false-negative test results). - Two-stage training: first supervised fine-tuning on synthetic Calgary-Cambridge interviews, then DAPO to optimize accuracy + tool quality + exam cost. - Achieves diagnostic performance comparable to larger models while keeping exam costs low. - Demonstrates that training on structured interview formats (Calgary-Cambridge) transfers to noisy, interactive diagnosis tasks.

arxiv cs.CL (NLP)

arXiv:2605.07058v1 Announce Type: new Abstract: Real-world clinical diagnosis is a complex process in which the doctor is required to obtain information from both interaction with the patient and conducting medical exams. Additionally, the doctor needs to adapt to different patient personas, as well as noisy and incomplete information that can happen at any time during the process. However, existing benchmarks for medical LLMs and methods for automatic diagnosis largely simplify this process by reducing it to single-turn question answering, noise-free conversations, or sequential exam making, etc., ignoring the interactive and uncertain nature of clinical diagnosis. In this paper, we aim to address this gap by formalizing clinical diagnosis as a Partially Observable Markov Decision Process (POMDP) with three action types: questioning the patient, ordering medical exams as tool calls, and issuing a diagnosis. We also introduce a systematic noise model comprising seven patient noise types and three exam noise types. Using our proposed environment, we train an effective diagnosis agent, \textbf{MedExAgent}, through a two-stage pipeline that first performs supervised finetuning on synthetic conversations structured after the Calgary-Cambridge model for clinical interviews, and then applies DAPO to optimize a composite reward capturing diagnostic accuracy, tool call quality, and exam cost including financial cost and patient discomfort. Through extensive experiments and ablation studies, we demonstrate that MedExAgent achieves diagnostic performance comparable to larger models while maintaining cost-efficient examination strategies.

Potential threat/caveat for clean result "Fine-tuning one persona on a two-marker chunk and another on the start marker plants the end marker at every donor answer's end, not chained to the start (LOW confidence)": this item discusses benchmark.

GSM-SEM: Benchmark and Framework for Generating Semantically Variant Augmentations

Jyotika Singh, Fang Tu, Aziza Mirzadova, Amit Agarwal, Hitesh Laxmichand Patel, Sandip Ghoshal, Miguel Ballesteros, Yassine Benajiba, Weiyi Sun, Graham Horwood, Sujith Ravi, Dan Roth · 2026-05-12

The authors introduce GSM-SEM, a framework for generating math-problem variants with high semantic diversity rather than just surface-level perturbations. Most robustness variants of GSM8K swap names or numbers but keep the underlying facts intact; GSM-SEM modifies entities, attributes, and relationships to produce problems that frequently alter the facts and require recomputation, while constraining generation to preserve the original answer and difficulty. Applying GSM-SEM to GSM8K, GSM-Symbolic, and GSM-Plus, they find consistent performance drops across 14 SOTA LLMs (28% average drop at maximum strictness), with larger declines when semantic perturbations combine with symbolic/plus variations. The framework generates fresh variants on each run, reducing reliance on static benchmarks and lowering memorization bias. **Main takeaways:** - GSM-SEM perturbs problem semantics (entities, attributes, relationships) rather than just surface features, often changing underlying facts while preserving the answer. - Generates fresh variants stochastically on each run, so it's not a fixed benchmark that models can memorize over time. - Applied to GSM8K, GSM-Symbolic, and GSM-Plus, producing three fully human-validated SEM datasets. - 14 SOTA LLMs show consistent performance drops (28% average at maximum strictness), especially when semantic + symbolic/plus perturbations combine. - Demonstrates applicability beyond math by applying GSM-SEM to BigBenchHard, LogicBench, and NLR-BIRD.

arxiv cs.CL (NLP)

arXiv:2605.07053v1 Announce Type: new Abstract: Benchmarks like GSM8K are popular measures of mathematical reasoning, but leaderboard gains can overstate true capability due to memorization of fixed test sets. Most robustness variants apply surface-level perturbations (paraphrases, renamings, number swaps, distractors) that largely preserve the underlying facts, and static releases can themselves become memorization targets over time. We introduce GSM-SEM, a reusable and stochastic framework for generating semantically diverse benchmark variants with substantially higher semantic variance than prior approaches. GSM-SEM perturbs problem statements by modifying entities, attributes, and/or relationships, frequently altering underlying facts and requiring models to recompute solutions under new conditions, while constraining generation to preserve the original calculations/answer and approximate problem difficulty. GSM-SEM generates fresh variants on each run without requiring re-annotation, reducing reliance on static public benchmarks for evaluation and thereby lowering the bias of memorization. We apply GSM-SEM on GSM8K and two existing variation suites (GSM-Symbolic and GSM-Plus), producing GSM8K-SEM, GSM-Symbolic-SEM, and GSM-Plus-SEM. Evaluating 14 SOTA LLMs, we observe consistent performance drops with larger decline when semantic perturbations are coupled with symbolic/plus variations (average drop rate 28% in maximum strictness configuration of GSM-SEM). We publicly release the three SEM variants as fully human-validated datasets. Finally, to demonstrate applicability beyond GSM-style math problems, we apply GSM-SEM to additional benchmarks including BigBenchHard, LogicBench, and NLR-BIRD.

Potential threat/caveat for clean result "Fine-tuning one persona on a two-marker chunk and another on the start marker plants the end marker at every donor answer's end, not chained to the start (LOW confidence)": this item discusses bias, robustness, evaluation, benchmark.

NSMQ Riddles: A Benchmark of Scientific and Mathematical Riddles for Quizzing Large Language Models

George Boateng, Naafi Ibrahim, Samuel John, Philemon Badu, Patrick Agyeman-Budu, Jonathan Mensah, Kevin Yeboah, William Edor, Andrew Mensa-Onumah, Nana Yeboah, Victor Wumbor-Apin Kumbol · 2026-05-12

The authors present NSMQ Riddles, a benchmark of scientific and mathematical riddles from Ghana's National Science and Maths Quiz, a live TV competition for high school students. Each riddle contains at least three clues (with earlier, vaguer clues fetching more points), and answers are numbers, words, or short phrases, enabling automatic evaluation. The dataset spans 11 years and 1,800 riddles covering biology, chemistry, physics, and math. State-of-the-art LLMs (GPT-4o, Gemini 2.0 Pro, Claude Opus 4, and open models like Kimi-K1.5 and DeepSeek-V3) performed worse than the best student contestants, even in high-reasoning settings, showing the dataset is challenging. **Main takeaways:** - 1,800 riddles from an 11-year archive of Ghana's National Science and Maths Quiz, each with 3+ clues of increasing specificity. - Answers are short (number, word, or phrase), making automatic evaluation straightforward. - SOTA closed and open LLMs (tested with high and low reasoning settings) perform worse than top student contestants. - Addresses underrepresentation of Global South datasets in LLM evaluation. - Contributes a challenging benchmark for scientific and mathematical reasoning beyond Western educational contexts.

arxiv cs.CL (NLP)

arXiv:2605.07051v1 Announce Type: new Abstract: Large Language Models (LLMs) have shown good performance on various science educational benchmarks, demonstrating their potential for use in science and mathematics education. Yet, LLMs tend to be evaluated on science and mathematical educational datasets from the Western world, with an underrepresentation of datasets from the Global South. Furthermore, they tend to have multiple-choice answer options that are trivial to evaluate. In this work, we present NSMQ Riddles, a novel benchmark of Scientific and Mathematical Riddles from Ghana's National Science and Maths Quiz (NSMQ) competition to evaluate LLMs. The NSMQ is an annual live TV competition for senior secondary school students in Ghana that brings together the smartest high school students in Ghana who compete in teams of 2 by answering questions in biology, chemistry, physics, and math over five rounds and five stages until a winning team is crowned for that year. NSMQ Riddles consists of 11 years of riddle questions (n=1.8K) from the 5th round, with each riddle containing a minimum of 3 clues. Students compete to be the first to guess the answer on any of the clues, with earlier clues being vague and also fetching more points. The answers are usually a number, word, or short phrase, allowing for automatic evaluation. We evaluated state-of-the-art models: closed (GPT-5.4, Gemini 3.1 Pro, Claude Opus 4.6) and open models (Kimi-K2.5, DeepSeek-V3.1, GPT-OSS-120B) with high and low reasoning settings. Our evaluation shows that the dataset is challenging even for state-of-the-art LLMs, which performed worse than the best student contestants. This work contributes a novel and challenging benchmark for scientific and mathematical reasoning from the Global South towards enabling a true global benchmarking of LLMs' capabilities for science and mathematics education.

Potential threat/caveat for clean result "Stretching turn count, completion length, or system-prompt length at train time fails to amplify marker uptake; the longest system prompt instead leaks across bystander personas (LOW confidence)": this item discusses evaluation, benchmark.

Cognitive Agent Compilation for Explicit Problem Solver Modeling

Hyeongdon Moon, Carolyn Ros\'e, John Stamper · 2026-05-12

The authors propose Cognitive Agent Compilation (CAC), a framework inspired by cognitive architectures that uses a strong "teacher" LLM to compile problem-solving knowledge into an explicit, inspectable "target agent." CAC separates (i) knowledge representation, (ii) problem-solving policy, and (iii) verification/update rules, making the agent's knowledge state and decisions transparent and editable—useful in educational settings where educators want to know what the system assumes the learner knows. They present an early proof-of-concept with small language models, surfacing design trade-offs between explicit control and scalable generalization, and position CAC as a step toward bounded-knowledge AI for education. **Main takeaways:** - CAC compiles problem-solving knowledge from a strong teacher LLM into an explicit, inspectable target agent with separated knowledge, policy, and update rules. - Goal is to make AI tutors' knowledge states transparent and editable for educators and learners. - Early proof-of-concept with small LMs highlights trade-offs between explicit control (inspectability, editability) and scalable generalization. - Inspired by cognitive architectures that use symbolic, inspectable knowledge representations. - Positions CAC as a building block for bounded-knowledge AI in educational applications.

arxiv cs.CL (NLP)

arXiv:2605.07040v1 Announce Type: new Abstract: Large language models (LLMs) are widely used for tutoring, feedback generation, and content creation, but their broad pretraining makes them hard to constrain and poor substitutes for controllable learners. Educational systems often require inspectable and editable knowledge states: educators want to know what a system assumes the learner knows, and learners benefit when the system can justify actions in terms of explicit skills, misconceptions, and strategies. Inspired by cognitive architectures, we propose Cognitive Agent Compilation (CAC), a framework that uses a strong teacher LLM to compile problem-solving knowledge into an explicit target agent. CAC separates (i) knowledge representation, (ii) problem-solving policy, and (iii) verification and update rules, with the goal of making bounded problem solving more inspectable and editable in educational settings. We present an early proof of concept implemented with Small Language Models that surfaces key design trade-offs, particularly between explicit control and scalable generalization, and positions CAC as an initial step toward bounded-knowledge AI for educational applications.

Towards Closing the Autoregressive Gap in Language Modeling via Entropy-Gated Continuous Bitstream Diffusion

Georgios Batzolis, Mark Girolami, Luca Ambrogioni · 2026-05-12

The authors propose a diffusion language model that represents text as a continuous diffusion process over fixed-width binary bitstreams, aiming to close the gap with autoregressive models. Each semantic token is represented as an analog bit sequence, and the model uses a matched-filter residual parameterization to separate contextual learning from the independent-bit posteriors. Crucially, they introduce a stochastic sampler that applies Langevin-type corrections gated by the entropy-rate profile—concentrating randomness in high-information regions while remaining nearly deterministic elsewhere. On LM1B, their 130M-parameter model reaches a generative perplexity of 59.76 (matching the autoregressive reference) using 256 function evaluations; on OpenWebText, it achieves 27.06 perplexity with 4× fewer steps than prior 1024-step baselines. Bitstream diffusion also removes the vocabulary-size scaling bottleneck, predicting O(log V) bitwise logits instead of O(V) token logits. **Main takeaways:** - Represents text as a diffusion process over continuous binary bitstreams (fixed-width analog bits) rather than discrete tokens. - Stochastic sampler gates Langevin corrections by entropy rate, concentrating randomness where information is high and being deterministic elsewhere. - On LM1B: 130M-param model hits generative perplexity 59.76 (matching autoregressive baseline) in 256 neural function evaluations. - On OpenWebText: achieves perplexity 27.06 with 4× fewer steps than prior 1024-step diffusion baselines. - Removes O(V) vocabulary scaling bottleneck by predicting O(log V) bitwise logits, reducing memory and increasing throughput.

arxiv cs.CL (NLP)

arXiv:2605.07013v1 Announce Type: new Abstract: Diffusion language models (DLMs) promise parallel, order-agnostic generation, but on standard benchmarks they have historically lagged behind autoregressive models in sample quality and diversity. Recent continuous flow and diffusion approaches over token embeddings have narrowed this gap, suggesting continuous state spaces are highly effective for language. In this work, we further close the autoregressive gap by modeling text as a continuous diffusion process over fixed-width binary bitstreams. Our approach represents semantic tokens as analog bit sequences and utilizes a matched-filter residual parameterization to isolate contextual learning from analytic independent-bit posteriors. Crucially, we adopt a stochastic sampler that applies Langevin-type corrections gated by the entropy-rate profile, automatically concentrating stochasticity in high-information regions while remaining nearly deterministic elsewhere. On the One Billion Word Benchmark (LM1B), our 130M-parameter bitstream model reaches a generative perplexity ($\GenPPL$) of $59.76$ at matched real-data entropy ($4.31$) using 256 neural function evaluations (NFEs), decisively outperforming prior DLM baselines and reaching the autoregressive reference. On OpenWebText (OWT), our stochastic sampler establishes a new continuous-DLM Pareto frontier, achieving $\GenPPL=27.06$ at an entropy of $5.26$ using $4\times$ fewer steps than previous 1024-NFE baselines. As an additional architectural benefit, bitstream diffusion removes the $\mathcal{O}(V)$ vocabulary scaling bottleneck shared by standard DLMs. By predicting $\mathcal{O}(\log V)$ bitwise logits via semantic bit-patching, our model yields a reduced memory footprint and higher throughput, demonstrating a scalable paradigm for language generation as vocabulary sizes grow.

Potential threat/caveat for clean result "Fine-tuning one persona on a two-marker chunk and another on the start marker plants the end marker at every donor answer's end, not chained to the start (LOW confidence)": this item discusses evaluation, benchmark.

Group of Skills: Group-Structured Skill Retrieval for Agent Skill Libraries

Kun Zeng, Yu Huo, Siyu Zhang, Zi Ye, Yuecheng Zhuo, Haoyue Liu, Yuquan Lu, Junhao Wen, Xiaoying Tang · 2026-05-12

The authors tackle a practical problem in agent systems: when an AI agent needs to retrieve skills from a large library, returning a flat list of individual skills forces the agent to figure out which one to run first, which are helpers, and how they fit together. Their "Group of Skills" method instead packages retrieved skills into a structured bundle with explicit labels—Start (the entry point), Support (helper skills), Check (preconditions), and Avoid (failure modes)—so the agent receives ready-to-execute context rather than raw pieces. Tests on two benchmarks (SkillsBench and ALFWorld) show this structured retrieval keeps important requirements visible even under tight memory budgets and often speeds up task completion. **Main takeaways:** - Traditional skill retrieval returns atomic skills or dependency bundles but leaves the agent to infer execution order and roles, which wastes inference cycles and risks missing preconditions. - GoSkills builds a typed graph of skills, clusters them around "anchor" skills, and outputs a fixed contract with four fields: Start, Support, Check, and Avoid. - The method works at inference time—no retraining of the agent, the skill payloads, or the execution environment. - Experiments show better coverage of requirements under memory constraints and improved task success and runtime compared to flat-list baselines. - The approach is modular: you get structured context without changing how skills are stored or how the agent executes them.

arxiv cs.CL (NLP)

arXiv:2605.06978v1 Announce Type: new Abstract: Skill-augmented agents increasingly rely on large reusable skill libraries, but retrieving relevant skills is not the same as presenting usable context. Existing methods typically return atomic skills or dependency-aware bundles whose internal roles remain implicit, leaving the agent to infer the execution entry point, support skills, visible requirements, and failure-avoidance guidance. We introduce Group of Skills (GoSkills), an inference-time group-structured retrieval method that changes the agent-facing retrieval object from a flat skill list to a compact, role-labeled execution context. GoSkills builds anchor-centered skill groups from a typed skill graph, expands support groups through a group graph, bottlenecks the selected group plan into a bounded set of atomic skill payloads, and renders a fixed execution contract with Start, Support, Check, and Avoid fields, without changing the downstream agent, skill payloads, or execution environment. Experiments on SkillsBench and ALFWorld show that GoSkills preserves visible-requirement coverage under a small skill budget, improves over flat skill-access baselines, and often improves reward and agent-only runtime relative to structural retrieval references.

Potential threat/caveat for clean result "Stretching turn count, completion length, or system-prompt length at train time fails to amplify marker uptake; the longest system prompt instead leaks across bystander personas (LOW confidence)": this item discusses failure.

MultiSoc-4D: A Benchmark for Diagnosing Instruction-Induced Label Collapse in Closed-Set LLM Annotation of Bengali Social Media

Souvik Pramanik, S. M. Riaz Rahman Antu, Shak Mohammad Abyad, Md. Ibrahim Khalil, Md. Shahriar Hussain · 2026-05-12

The authors built MultiSoc-4D, a 58,000-comment Bengali social-media dataset annotated for category, sentiment, hate speech, and sarcasm, and used it to diagnose a systematic LLM annotation failure they call "instruction-induced label collapse." When asked to label closed-set categories (e.g., hateful/not hateful), ChatGPT, Gemini, Claude, and Grok all showed strong bias toward safe fallback labels ("Other," "Neutral," "No"), missing 79% of hateful content and 75% of sarcasm compared to human-calibrated labels. Even though the models agreed with each other at high rates, Fleiss' kappa was near zero for sarcasm, revealing what the authors call a "label agreement illusion"—models converge on the wrong answer together. **Main takeaways:** - LLMs systematically prefer neutral or "Other" labels in closed-set annotation tasks, especially for minority categories like hate speech and sarcasm. - High inter-LLM agreement doesn't guarantee quality: the models can all collapse to the same wrong default, producing near-zero kappa despite surface consensus. - The effect persists across 40+ LLMs of different architectures, suggesting it's a widespread training-pipeline issue rather than model-specific. - The dataset is released as a diagnostic benchmark for annotation bias in low-resource (Bengali) NLP. - The bias propagates downstream: models trained on LLM-annotated data inherit the label collapse.

arxiv cs.CL (NLP)

arXiv:2605.06940v1 Announce Type: new Abstract: Annotation automation via Large Language Models (LLMs) is the core approach for scaling NLP datasets; however, LLM behavior with respect to closed-set instructions in low-resource languages has not been well studied. We present MultiSoc-4D, a Bengali social media dataset benchmark, which contains 58K+ social media comments from six sources annotated along four dimensions: category, sentiment, hate speech, and sarcasm. By employing a structured pipeline where ChatGPT, Gemini, Claude, and Grok individually annotate separate partitions, while sharing a common validation set of 20%, we diagnose LLM behavior systematically. We discover a prevalent phenomenon called "instruction-induced label collapse", wherein LLMs show a systematic preference towards fallback labels (Other, Neutral, No), leading to high agreement rates but under-detection of minority categories. For example, we find that LLMs failed to detect 79% and 75% of instances with hateful and sarcastic content compared to a human-calibrated reference. Furthermore, we prove that it represents a "label agreement illusion", statistically validated via almost null Fleiss' Kappa ($\kappa \approx -0.001$) on sarcasm detection. Across 40+ LLMs, we benchmark this annotation bias propagation within the training pipeline, regardless of architectural differences. We release MultiSoc-4D as a diagnostic benchmark for annotation biases in Bengali NLP.

Potential threat/caveat for clean result "Fine-tuning one persona on a two-marker chunk and another on the start marker plants the end marker at every donor answer's end, not chained to the start (LOW confidence)": this item discusses bias, benchmark.

Can LLMs Take Retrieved Information with a Grain of Salt?

Behzad Shayegh, Mohamed Osama Ahmed, Fred Tung, Leo Feng · 2026-05-12

The authors test whether LLMs can adjust their answers to match how certain the retrieved context is—e.g., hedging when told a fact is "uncertain" or being confident when it's "certain." They find systematic failures across eight models: LLMs forget their own prior knowledge after seeing uncertain context, misinterpret certainty cues, and overtrust complex contexts. To fix this without retraining, they propose a three-part interaction strategy: remind the model of its prior knowledge, explicitly recalibrate the certainty level in the prompt, and simplify the context. This reduces "obedience errors" (mismatches between context certainty and response confidence) by 25% on average. **Main takeaways:** - LLMs struggle to recall prior knowledge after observing an uncertain context, even when the context explicitly says "this might be wrong." - They misinterpret expressed certainties—e.g., treating "possibly" the same as "definitely." - Complex or verbose contexts are overtrusted regardless of stated certainty level. - A prompt-engineering fix (prior reminder + certainty recalibration + simplification) cuts obedience errors by ~25% without model changes. - The evaluation metric, "context-certainty obedience," measures how well response confidence tracks stated context certainty, which matters in high-stakes domains like medicine and finance.

arxiv cs.CL (NLP)

arXiv:2605.06919v1 Announce Type: new Abstract: Large language models have demonstrated impressive retrieval-augmented capabilities. However, a crucial area remains underexplored: their ability to appropriately adapt responses to the certainty of the retrieved information. It is a limitation with real consequences in high-stakes domains like medicine and finance. We evaluate eight LLMs on their context-certainty obedience, measuring how well they adjust responses to match expressed context certainty. Our analysis reveals systematic limitations: LLMs struggle to recall prior knowledge after observing an uncertain context, misinterpret expressed certainties, and overtrust complex contexts. To address these, we propose an interaction strategy combining prior reminders, certainty recalibration, and context simplification. This approach reduces obedience errors by 25% on average, without modifying model weights, demonstrating the efficacy of interaction design in enhancing LLM reliability. Our contributions include a principled evaluation metric, empirical insights into LLMs' uncertainty handling, and a portable strategy to improve context-certainty obedience across diverse LLMs.

Potential threat/caveat for clean result "Fine-tuning one persona on a two-marker chunk and another on the start marker plants the end marker at every donor answer's end, not chained to the start (LOW confidence)": this item discusses limitation, limitations, evaluation.

MELD: Multi-Task Equilibrated Learning Detector for AI-Generated Text

Chenjun Li, Cheng Wan, Johannes C. Paetzold · 2026-05-12

The authors built MELD, a detector for AI-generated text that goes beyond binary human-vs-AI classification by adding auxiliary tasks: predicting which model family generated the text, what kind of adversarial attack was applied, and what domain it came from. They train all four tasks together on a shared encoder, balancing the losses with learned uncertainty weights, and use teacher-student distillation (clean teacher, attack-augmented student) plus a hard-negative ranking loss to widen the score gap between AI text and confusable human text. At inference, the auxiliary heads are discarded, so MELD costs the same as a standard detector. On the public RAID leaderboard, MELD is the strongest open-source detector and competes with commercial models, especially under attack and at low false-positive rates. **Main takeaways:** - Most AI-text detectors optimize only binary human/AI classification, so the representation doesn't learn fine-grained structure (generator, attack type, domain) once the binary task saturates. - MELD adds three auxiliary heads (generator family, attack type, source domain) to a shared encoder, forcing the representation to capture richer signal. - Losses are balanced with learned homoscedastic uncertainty weights (i.e., the model learns how much to weight each task). - Teacher-student distillation with attack augmentation on the student improves robustness; hard-negative ranking enlarges the margin between AI and hard-to-distinguish human text. - MELD tops open-source detectors on the RAID leaderboard and matches or beats supervised baselines on held-out benchmarks, especially at low false-positive rates and under adversarial rewrites.

arxiv cs.CL (NLP)

arXiv:2605.06903v1 Announce Type: new Abstract: Large language models are now embedded in everyday writing workflows, making reliable AI-generated text detection important for academic integrity, content moderation, and provenance tracking. In practice, however, a detector must do more than achieve high aggregate AUROC on clean, in-distribution human and AI text: it should remain robust to attacks and adversarial rewrites, transfer to unseen generators and domains, and operate at low false-positive rates (FPR). Most existing detectors optimize a single AI/Human objective, giving the representation little incentive to learn generator, attack, or domain structure once the binary task saturates. We introduce MELD (Multi-Task Equilibrated Learning Detector), a deployable detector for AI-generated text that enriches binary detection with auxiliary supervision. MELD attaches generator-family, attack-type, and source-domain heads to a shared encoder, and balances the four losses with learned homoscedastic uncertainty weights. To improve robustness, an EMA teacher predicts on clean inputs while an attack-augmented student is distilled toward the teacher. MELD further uses a hard-negative pairwise ranking loss to enlarge the score margin between AI-generated texts and the most confusable human texts. At inference, all auxiliary heads are discarded, giving MELD the same interface and cost as a standard detector. On the public RAID leaderboard, MELD is the strongest open-source detector and is competitive with leading commercial models, especially under attack and at low FPR. Across standard held-out benchmarks, MELD matches or outperforms supervised baselines. We further introduce MELD-eval, a held-out evaluation pool built from recent chat models released by four major LLM providers. Without additional finetuning, MELD achieves 99.9% TPR at 1% FPR on MELD-eval, while many baselines degrade sharply.

Potential threat/caveat for clean result "Fine-tuning one persona on a two-marker chunk and another on the start marker plants the end marker at every donor answer's end, not chained to the start (LOW confidence)": this item discusses negative, robustness, adversarial, evaluation, benchmark.

Reflections and New Directions for Human-Centered Large Language Models

Caleb Ziems, Dora Zhao, Rose E. Wang, Matthew J\"orke, Ahmad Rushdi, Advit Deepak, Sunny Yu, Anshika Agarwal, Harshvardhan Agarwal, Gabriela Aranguiz-Dias, Aditri Bhagirath, Justine Breuch, Huanxing Chen, Ruishi Chen, Sarah Chen, Haocheng Fan, William Fang, Cat Gonzales Fergesen, Daniel Frees, Tian Gao, Ziqing Huang, Vishal Jain, Yucheng Jiang, Kirill Kalinin, Su Doga Karaca, Arpandeep Khatua, Teland La, Isabelle Levent, Miranda Li, Xinling Li, Yongce Li, Angela Liu, Minsik Oh, Nathan J. Paek, Anthony Qin, Emily Redmond, Michael J. Ryan, Aadesh Salecha, Xiaoxian Shen, Pranava Singhal, Shashanka Subrahmanya, Mei Tan, Irawadee Thawornbut, Michelle Vinocour, Xiaoyue Wang, Zheng Wang, Henry Jin Weng, Pawan Wirawarn, Shirley Wu, Sophie Wu, Yichen Xie, Patrick Ye, Sean Zhang, Yutong Zhang, Cathy Zhou, Yiling Zhao, James Landay, Diyi Yang · 2026-05-12

This position paper from a large interdisciplinary team argues that "human-centered" LLM development should be rigorous and integrated at every pipeline stage—data sourcing, training, evaluation, deployment—rather than bolted on during post-training or alignment. The authors synthesize perspectives from NLP, human-computer interaction, and responsible AI to propose a framework for Human-Centered LLMs (HCLLMs), emphasizing user needs, values, and real-world context alongside technical benchmarks. The paper offers stage-by-stage recommendations and closes with a case study on the future of work with LLMs. **Main takeaways:** - Most current "human-centered" efforts happen late (post-training RLHF or red-teaming) rather than being designed in from the start. - The authors propose integrating human priorities—ethics, preferences, values, accessibility—at every stage: system design, data curation, model training, evaluation, and deployment. - The framework draws on HCI (user experience, participatory design) and responsible AI (fairness, transparency, safety) alongside NLP technical goals. - The paper is a roadmap and call to action rather than an empirical study; it offers recommendations and a case study on work contexts. - Key insight: technical capability alone doesn't guarantee beneficial or safe real-world impact; centering human concerns requires intentional design choices throughout the pipeline.

arxiv cs.CL (NLP)

arXiv:2605.06901v1 Announce Type: new Abstract: Large Language Models (LLMs) are increasingly shaping the private and professional lives of users, with numerous applications in business, education, finance, healthcare, law, and science. With this rise in global influence comes greater urgency to build, evaluate, and deploy these systems in a manner that prioritizes not only technical capabilities but also human priorities. This work presents a framework for developing Human-Centered Large Language Models (HCLLMs), which integrates perspectives from Natural Language Processing (NLP), Human-Computer Interaction (HCI), and responsible AI. Considering the ethics, economics, and technical objectives of language modeling, we argue that model developers need to address human concerns, preferences, values, and goals, not only during a cursory post-training stage, but rather with rigor and care at every stage of the pipeline. This paper offers human-centered insights and recommendations for developers at each stage, from system design to data sourcing, model training, evaluation, and responsible deployment. Then we conclude with a case study, applying these insights to understand the future of work with HCLLMs.

Potential threat/caveat for clean result "Fine-tuning one persona on a two-marker chunk and another on the start marker plants the end marker at every donor answer's end, not chained to the start (LOW confidence)": this item discusses evaluation.

MIST: Multimodal Interactive Speech-based Tool-calling Conversational Assistants for Smart Homes

Maximillian Chen, Xuanming Zhang, Michael Peng, Zhou Yu, Alexandros Papangelis, Yohan Jo · 2026-05-12

The authors introduce MIST, a synthetic benchmark for voice-driven smart-home assistants that combines speech input, tool-calling over IoT devices, and mixed-initiative multi-turn dialogue. The task requires models to generate code that respects spatiotemporal constraints (e.g., "turn off the lights in the kitchen but only if no one is there"), track dynamic device state across turns, and handle interruptions or clarifications from the user. Benchmarking open- and closed-weight multimodal LLMs reveals a large gap: even frontier closed models have substantial room for improvement. The dataset and generation framework are released to support research on voice assistants that reason about the physical world. **Main takeaways:** - MIST combines speech inputs, multi-turn dialogue, tool-calling (code generation for IoT devices), and spatiotemporal reasoning in one benchmark. - The task is synthetic but designed to reflect real smart-home complexity: dynamic state, mixed initiative (user can interrupt or clarify), and physical-world constraints. - Open-weight multimodal LLMs lag far behind closed-weight models; even the best closed models have significant headroom. - The authors release both the dataset and an extensible data-generation framework so others can create similar benchmarks for related domains. - Key challenge: modeling physical-world constraints ("Is anyone in the room?") alongside traditional NLP reasoning.

arxiv cs.CL (NLP)

arXiv:2605.06897v1 Announce Type: new Abstract: The rise of Internet of Things (IoT) devices in the physical world necessitates voice-based interfaces capable of handling complex user experiences. While modern Large Language Models (LLMs) already demonstrate strong tool-usage capabilities, modeling real-world IoT devices presents a difficult, understudied challenge which combines modeling spatiotemporal constraints with speech inputs, dynamic state tracking, and mixed-initiative interaction patterns. We introduce MIST (the Multimodal Interactive Speech-based Tool-calling Dataset), a synthetic multi-turn, voice-driven code generation task that operates over IoT devices. We find that there is a significant gap between open- and closed-weight multimodal LLMs on MIST, and that even frontier closed-weight LLMs have substantial headroom. We release MIST and an extensible data generation framework to build related datasets in order to facilitate research on mixed-initiative voice assistants which reason about physical world constraints.

TajPersLexon: A Tajik-Persian Lexical Resource and Hybrid Model for Cross-Script Low-Resource NLP

Mullosharaf K. Arabov · 2026-05-12

The author presents TajPersLexon, a curated 40,112-pair Tajik-Persian parallel lexicon for cross-script lexical tasks (retrieval, transliteration, alignment) in a low-resource setting. The benchmark compares three families of methods—lightweight hybrid pipelines, neural sequence-to-sequence, and retrieval—on CPU only. Neural and retrieval baselines hit 98–99% top-1 accuracy, essentially solving exact lexical matching, but large multilingual sentence transformers fail. The hybrid model offers a practical accuracy-efficiency trade-off, achieving 96.4% accuracy on an OCR post-correction task, and is interpretable and fast. All experiments use fixed random seeds for full reproducibility; dataset, code, and models will be released. **Main takeaways:** - TajPersLexon is a 40k-pair Tajik-Persian lexicon for cross-script word/short-phrase matching (Cyrillic Tajik ↔ Persian script). - Neural seq2seq and retrieval methods nearly solve the task (98–99% top-1 accuracy) on exact lexical matching. - Multilingual sentence transformers fail despite being large, showing that sentence-level embeddings don't transfer to exact lexical tasks. - The lightweight hybrid model (rule-based + small neural components) is interpretable, CPU-friendly, and achieves 96.4% on real-world OCR correction. - All code, data, and models will be open-sourced with fixed seeds for reproducibility.

arxiv cs.CL (NLP)

arXiv:2605.06886v1 Announce Type: new Abstract: This work introduces TajPersLexon, a curated Tajik--Persian parallel lexical resource of 40,112 word and short-phrase pairs for cross-script lexical retrieval, transliteration, and alignment in low-resource settings. We conduct a comprehensive CPU-only benchmark comparing three methodological families: (i) a lightweight hybrid pipeline, (ii) neural sequence-to-sequence models, and (iii) retrieval methods. Our evaluation establishes that the task is essentially solvable, with neural and retrieval baselines achieving 98-99% top-1 accuracy. Crucially, we demonstrate that while large multilingual sentence transformers fail on this exact lexical matching, our interpretable hybrid model offers a favorable accuracy-efficiency trade-off for practical applications, achieving 96.4% accuracy in an OCR post-correction task. All experiments use fixed random seeds for full reproducibility. The dataset, code, and models will be publicly released.

Potential threat/caveat for clean result "Fine-tuning one persona on a two-marker chunk and another on the start marker plants the end marker at every donor answer's end, not chained to the start (LOW confidence)": this item discusses evaluation, benchmark.

IntentGrasp: A Comprehensive Benchmark for Intent Understanding

Yuwei Yin, Chuyuan Li, Giuseppe Carenini · 2026-05-12

The authors built IntentGrasp, a large-scale benchmark for evaluating how well LLMs understand intent in speech, conversation, and writing. Drawn from 49 diverse, open-licensed corpora across 12 domains, the benchmark includes 262k training instances, a 12.9k All Set, and a harder 470-case Gem Set. Testing 20 LLMs (including GPT-5.4, Gemini-3.1-Pro, Claude-Opus-4.7) reveals poor performance: below 60% on All Set, below 25% on Gem Set, with 17 of 20 models worse than random guessing (15.2%) on Gem Set, while estimated human performance is ~81%. The authors propose Intentional Fine-Tuning (IFT)—fine-tuning on the training set—which boosts F1 by 30+ points on All Set and 20+ on Gem Set, with strong cross-domain generalization in leave-one-domain-out experiments. **Main takeaways:** - Intent understanding is crucial for helpful assistants but remains unsolved: frontier LLMs score below 60% on All Set and below 25% on the harder Gem Set. - 17 of 20 tested models perform worse than random guessing on Gem Set; estimated human performance is ~81%, showing huge headroom. - IntentGrasp unifies 49 corpora across 12 domains into a single benchmark with a large training set (262k) and two test sets (All and Gem). - Intentional Fine-Tuning (IFT)—simply fine-tuning on the training set—yields 30+ point F1 gains on All Set and 20+ on Gem Set. - Leave-one-domain-out experiments confirm IFT generalizes well across domains, suggesting it's a promising path to more capable and safer assistants.

arxiv cs.CL (NLP)

arXiv:2605.06832v1 Announce Type: new Abstract: Accurately understanding the intent behind speech, conversation, and writing is crucial to the development of helpful Large Language Model (LLM) assistants. This paper introduces IntentGrasp, a comprehensive benchmark for evaluating the intent understanding capability of LLMs. Derived from 49 high-quality, open-licensed corpora spanning 12 diverse domains, IntentGrasp is constructed through source datasets curation, intent label contextualization, and task format unification. IntentGrasp contains a large-scale training set of 262,759 instances and two evaluation sets: an All Set of 12,909 test cases and a more balanced and challenging Gem Set of 470 cases. Extensive evaluations on 20 LLMs across 7 families (including frontier models such as GPT-5.4, Gemini-3.1-Pro, and Claude-Opus-4.7) demonstrate unsatisfactory performance, with scores below 60% on All Set and below 25% on Gem set. Notably, 17 out of 20 tested models perform worse than a random-guess baseline (15.2%) on Gem Set, while the estimated human performance is ~81.1%, showing substantial room for improvement. To enhance such ability, this paper proposes Intentional Fine-Tuning (IFT), which fine-tunes the models on the training set in IntentGrasp, yielding significant gains of 30+ F1 points on All Set and 20+ points on Gem Set. Tellingly, the leave-one-domain-out (Lodo) experiments further demonstrate the strong cross-domain generalizability of IFT, verifying that it is a promising approach to substantially enhancing the intent understanding of LLMs. Overall, by benchmarking and boosting intent understanding ability, this study sheds light on a promising path towards more intentional, capable, and safe AI assistants for human benefits and social good.

Potential threat/caveat for clean result "Fine-tuning one persona on a two-marker chunk and another on the start marker plants the end marker at every donor answer's end, not chained to the start (LOW confidence)": this item discusses evaluation, benchmark.

VITA-QinYu: Expressive Spoken Language Model for Role-Playing and Singing

Jiacheng Xu, Heting Gao, Liufei Xie, Zhenchuan Yang, Lijiang Li, Yiting Chen, Bin Zhang, Meng Chen, Chaoyu Fu, Weifeng Zhao, Wenjiang Zhou · 2026-05-12

The authors built VITA-QinYu, a spoken language model that can do more than just chat — it can role-play different personalities and even sing. The model uses separate tokens for text and audio (with multiple codebooks for richer sound representation) and was trained on nearly 16,000 hours of conversation, role-playing, and singing data. It beats other spoken models both on objective role-playing tasks and on singing quality, while also improving conversational accuracy. **Main takeaways:** - First end-to-end spoken model that handles natural conversation, role-playing (e.g., speaking in a comforting tone), and singing generation in one system - Uses a hybrid design that keeps text and audio modalities separate to avoid them interfering with each other, while using multi-codebook tokens to capture expressive vocal features - Outperforms peer models by 7 percentage points on role-playing benchmarks and by 0.13 MOS points (on a 5-point scale) for singing - Also achieves state-of-the-art on conversational benchmarks, beating prior models by 1.4–5 percentage points - Code, models, and a streaming full-duplex demo are open-sourced

arxiv cs.CL (NLP)

arXiv:2605.06765v1 Announce Type: new Abstract: Human speech conveys expressiveness beyond linguistic content, including personality, mood, or performance elements, such as a comforting tone or humming a song, which we formalize as role-playing and singing. We present VITA-QinYu, the first expressive end-to-end (E2E) spoken language model (SLM) that goes beyond natural conversation to support both role-playing and singing generation. VITA-QinYu adopts a hybrid speech-text paradigm that extends interleaved text-audio modeling with multi-codebook audio tokens, a design enabling richer paralinguistic representation while preserving a clear separation between modalities to avoid interference. We further develop a comprehensive data generation pipeline to synthesize a total of 15.8K hours of natural conversation, role-playing, and singing data for training. VITA-QinYu demonstrates superior expressiveness, outperforming peer SLMs by 7 percentage points on objective role-playing benchmarks, and surpassing peer models by 0.13 points on a 5-point MOS scale for singing. Simultaneously, it achieves state-of-the-art conversational accuracy and fluency, exceeding prior SLMs by 1.38 and 4.98 percentage points on the C3 and URO benchmarks, respectively. We open-source our code and models and provide an easy-to-use demo with full-stack support for streaming and full-duplex interaction.

Potential threat/caveat for clean result "Fine-tuning one persona on a two-marker chunk and another on the start marker plants the end marker at every donor answer's end, not chained to the start (LOW confidence)": this item discusses benchmark.

Domain-level metacognitive monitoring in frontier LLMs: A 33-model atlas

Jon-Paul Cacioli · 2026-05-12

This paper tested how well 33 frontier language models know when they're right or wrong (metacognition) across six different knowledge domains from the MMLU benchmark. The author gave models 1,500 questions, asked them to rate their confidence (0-100), and measured whether high confidence actually predicted correct answers. Every model showed big variation across domains — models were consistently good at monitoring their accuracy on applied/professional knowledge but struggled with formal reasoning and natural science. **Main takeaways:** - All models with above-chance metacognition showed significant domain-to-domain variation that aggregate scores hide - Applied/Professional knowledge was easiest to monitor (mean AUROC = 0.742, top-2 in 21/33 models); Formal Reasoning and Natural Science were hardest (bottom-2 in 27/33 models) - Some model families (Anthropic, Gemini, Qwen) show consistent within-family patterns in which domains are hard vs. easy, while others (DeepSeek, Gemma, OpenAI) don't - Three models that failed on binary yes/no confidence probes worked fine when asked for 0-100 scores, showing that the prompt format matters a lot - Gemma 4 31B showed a massive +0.202 AUROC improvement over Gemma 3 27B, suggesting newer models are getting better at knowing what they know

arxiv cs.CL (NLP)

arXiv:2605.06673v1 Announce Type: new Abstract: Aggregate metacognitive quality scores mask within-model variation across MMLU benchmark domains. We administered 1,500 MMLU items (250 per domain, under an a priori six-domain grouping) to 33 frontier LLMs from eight model families and computed Type-2 AUROC per model-domain cell using verbalized confidence (0-100). Total observations: 47,151. Every model with above-chance aggregate monitoring showed non-trivial domain-level variation. Applied/Professional knowledge was reliably the easiest benchmark domain to monitor (mean AUROC = .742, ranked top-2 in 21 of 33 models); Formal Reasoning and Natural Science were reliably the hardest (one of the two ranked bottom-2 in 27 of 33 models). The three middle domains were statistically indistinguishable (Kendall's W = .164). A subject-level coherence analysis (within-domain similarity ratio = 0.95) confirms the six-domain grouping is a pragmatic benchmark taxonomy, not a validated latent construct. Within-family profile-shape clustering is significant for Anthropic, Google-Gemini, and Qwen (permutation p < .0001) but not DeepSeek, Google-Gemma, or OpenAI. Gemma 4 31B showed a +.202 AUROC improvement over Gemma 3 27B. Three models classified Invalid on binary KEEP/WITHDRAW probes produced normal profiles under verbalized confidence, confirming probe-format specificity. Bootstrap 95% CIs on 198 cells have median width .199. Split-half aggregate stability r = .893; profile-level split-half is weaker (grand median r = .184). These results show stable benchmark-domain variation obscured by aggregate metrics, and support benchmark-stage domain screening as a step before deployment in specific application areas.

Potential threat/caveat for clean result "A pretraining-data-poisoned Qwen3-4B backdoor only fires on the exact trigger tokens — paraphrases don't activate it, and base-model similarity to the trigger doesn't predict which inputs fire (MODERATE confidence)": this item discusses benchmark.

A Theory of Online Learning with Autoregressive Chain-of-Thought Reasoning

Ilan Doron-Arad, Idan Mehalel, Elchanan Mossel · 2026-05-11

The authors study how learnable autoregressive next-token generation is, framed as an online learning problem where you're trying to predict the final token after M steps of chain-of-thought. They compare two feedback regimes: end-to-end (you only see the final output) versus chain-of-thought (you see the entire intermediate trajectory). The key result is that seeing the intermediate reasoning tokens eliminates dependence on the generation horizon M, while end-to-end feedback can require logarithmic mistake growth. **Main takeaways:** - In the end-to-end setting (only final token visible), the number of mistakes needed to learn grows between constant and logarithmic in M (the number of autoregressive steps), and this logarithmic ceiling is unavoidable. - In the chain-of-thought setting (full trajectory visible), the mistake bound becomes independent of M entirely — seeing intermediate tokens is a huge advantage. - This echoes prior statistical learning results but at a different scale, showing the online theory has its own qualitative structure. - For specific function classes like linear thresholds, they prove tight bounds on how many mistakes are needed.

arxiv cs.LG (Machine Learning)

arXiv:2605.06819v1 Announce Type: new Abstract: Autoregressive generation lies at the heart of the mechanism of large language models. It can be viewed as the repeated application of a next-token generator: starting from an input string (prompt), the generator is applied for $M$ steps, and the last generated token is taken as the final output. [Joshi et al., 2025] proposed a PAC model for studying the learnability of the input-output maps arising from this process. We develop an online analogue of this framework, focusing on the mistake bound of learning the final output induced by an unknown next-token generator. We distinguish between two forms of feedback. In the End-to-End model, after each round the learner observes only the final token produced after $M$ autoregressive steps. In the Chain-of-Thought model, the learner is additionally shown the entire $M$-step trajectory. Our goal is to understand how the optimal mistake bound depends on the generation horizon $M$, and to what extent observing intermediate tokens can reduce this dependence. Our main results show that the online theory of autoregressive learning exhibits a qualitative picture analogous to the statistical one found by [Hanneke et al., 2026], but with a different scale of dependence on the generation horizon. In the End-to-End model, we prove a taxonomy of possible mistake-bound growth rates in the generation horizon $M$: essentially any rate between constant and logarithmic can arise. We further show that this logarithmic ceiling is unavoidable. In the Chain-of-Thought model, we show that access to the full generated trajectory eliminates the dependence on $M$ altogether. We also analyze autoregressive linear threshold classes, and prove optimal mistake bounds, as well as a new lower bound for the statistical setting. Along the way, our results resolve several questions left open by [Joshi et al., 2025].

From Model to Data (M2D): Shifting Complexity from GNNs to Graphs for Transparent Graph Learning

Debolina Halder Lina, Arlei Silva · 2026-05-11

The authors propose a new framework for making graph neural networks (GNNs) more transparent by moving complexity out of the model architecture and into the data itself. They "distill" a complex teacher GNN into an augmented graph with enriched features and structure, so that a simple student model can match the teacher's performance on this new graph. This lets humans inspect what the complex model was doing by looking at the augmented data rather than reverse-engineering opaque model internals. **Main takeaways:** - Standard GNN explainability methods highlight important nodes or edges for individual predictions, but don't explain why one architecture outperforms another. - Model-to-Data (M2D) distillation transfers architectural advantages (like attention mechanisms or fairness constraints) into visible graph structure and features. - A simple student GNN trained on the M2D-augmented graph matches the teacher's performance, making the learned behavior interpretable. - The approach reveals mechanisms like attention-based aggregation by materializing them as changes in the graph topology or features.

arxiv cs.LG (Machine Learning)

arXiv:2605.06814v1 Announce Type: new Abstract: Graph Neural Networks (GNNs) achieve high performance but can be opaque to humans, making it difficult to understand and compare the many proposed architectures. While existing explainability methods attribute individual predictions to nodes, edges, or features, they do not provide architectural transparency or explain the fundamental performance gap between simple and more complex models. To address this limitation, we introduce Model-to-Data (M2D) distillation, a new framework that increases transparency by transferring model complexity into the data space. M2D distills the teacher model into an augmented graph with enriched features and structure, enabling a simple student to match the teacher's performance. By materializing model behavior in the data, our approach allows humans to inspect architectural advantages directly. We show that M2D reveals underlying mechanisms such as fairness objectives and attention-based aggregation in an interpretable way, enhancing GNN transparency while preserving performance.

MIND: Monge Inception Distance for Generative Models Evaluation

Quentin Berthet, Yu-Han Wu, Clement Crepy, Romuald Elie, Klaus Greff, Michael Eli Sander · 2026-05-11

The authors introduce MIND (Monge Inception Distance), a new metric for evaluating generative models that fixes major problems with the widely used FID (Fréchet Inception Distance). MIND uses sliced Wasserstein distance—averaging many one-dimensional optimal transport problems—instead of comparing high-dimensional Gaussian statistics, which makes it much more sample-efficient, faster to compute, and robust to adversarial attacks like moment-matching. **Main takeaways:** - FID estimates high-dimensional means and covariances, which requires tons of samples and is fragile; MIND avoids this by using sliced optimal transport (sorting in 1D). - MIND with 5,000 samples performs as well as FID with 50,000 samples, a 10× improvement in sample efficiency. - MIND is two orders of magnitude faster to compute than FID. - Even at 1,000–2,000 samples, MIND remains highly informative for rapid model iteration, and it correlates well with FID while being more discriminative and attack-resistant.

arxiv cs.LG (Machine Learning)

arXiv:2605.06797v1 Announce Type: new Abstract: We propose the Monge Inception Distance (MIND), a metric for evaluating generative models that addresses key limitations of the widely adopted Fr\'echet Inception Distance (FID). The MIND metric leverages the sliced Wasserstein distance to compare distributions by averaging one-dimensional optimal transport distances, efficiently computed via sorting. This approach circumvents the estimation of high-dimensional means and covariance matrices, which underlie FID's poor sample complexity and vulnerability to adversarial attacks. We empirically demonstrate three primary advantages: (i) it is more sample-efficient by one order of magnitude, (ii) it is faster to compute by two orders of magnitude, (iii) it is more robust to adversarial attacks such as moment-matching. We show that MIND with 5k samples can replace the evaluation performance of FID with 50k samples, providing high correlation with this standard benchmark and superior discriminative performance. We further demonstrate that even smaller sample sizes (e.g., 1k or 2k) remain highly informative for rapid model iteration.

Conformal Agent Error Attribution

Naihe Feng, Yi Sui, Shiyi Hou, Ga Wu, Jesse C. Cresswell · 2026-05-11

The authors tackle error attribution in multi-agent systems (MAS) built from large language models, where failures leave long interaction traces and it's hard to pinpoint which step went wrong. They use conformal prediction (a framework that gives finite-sample, distribution-free coverage guarantees) to predict contiguous subsequences of the agent trajectory that likely contain the error, enabling automated rollback and recovery. **Main takeaways:** - Standard conformal prediction doesn't handle sequential data well; the authors introduce filtration-based conformal prediction tailored to agent trajectories. - Their method predicts contiguous sequence intervals (not arbitrary sets), which makes recovery practical—you can roll back to a specific earlier state. - The approach is model-agnostic and provides principled uncertainty quantification for error localization. - Experiments show errors can be precisely isolated, and agents can use the prediction sets to roll back and correct their own mistakes.

arxiv cs.LG (Machine Learning)

arXiv:2605.06788v1 Announce Type: new Abstract: When multi-agent systems (MAS) fail, identifying where the decisive error occurred is the first step for automated recovery to an earlier state. Error attribution remains a fundamental challenge due to the long interaction traces that large language model-based MAS generate. This paper presents a framework for error attribution based on conformal prediction (CP) which provides finite-sample, distribution-free coverage guarantees. We introduce new algorithms for filtration-based CP designed for sequential data such as agent trajectories. Unlike existing CP algorithms, our approach predicts sets that are contiguous sequences to enable efficient recovery and debugging. We verify our theoretical guarantees on a variety of agents and datasets, show that errors can be precisely isolated, then use prediction sets to rollback MAS to correct their own errors. Our overall approach is model-agnostic, and offers a principled uncertainty layer for MAS error attribution. We release code at https://github.com/layer6ai-labs/conformal-agent-error-attribution.

Distributional Process Reward Models: Calibrated Prediction of Future Rewards via Conditional Optimal Transport

Rachel Ma, Dylan Hadfield-Menell, Kristjan Greenewald · 2026-05-11

The authors propose using conditional optimal transport to calibrate process reward models (PRMs) used in inference-time scaling, which currently overestimate success probabilities and are poorly calibrated. They adapt a conditional optimal transport method to learn a monotonic conditional quantile function over PRM scores, giving well-calibrated confidence bounds at any desired level, and plug this into adaptive scaling frameworks that use PRMs to decide how much compute to spend. **Main takeaways:** - Process reward models (PRMs) guide search and scaling at inference time, but they often give overconfident probability estimates. - The authors use conditional optimal transport to map PRM hidden states to calibrated quantile estimates of success probability, preserving monotonicity. - This yields valid confidence intervals and integrates into instance-adaptive scaling (spending more compute when the model is uncertain). - On math reasoning benchmarks (MATH-500, AIME), the method improves calibration substantially and often improves downstream Best-of-N performance over uncalibrated PRMs.

arxiv cs.LG (Machine Learning)

arXiv:2605.06785v1 Announce Type: new Abstract: Inference-time scaling methods rely on Process Reward Models (PRMs), which are often poorly calibrated and overestimate success probabilities. We propose, to our knowledge, the first use of conditional optimal transport for calibrating PRMs, modifying conditional OT (CondOT) map learning \cite{bunne2022supervised} to estimate a monotonic conditional quantile function over success probabilities estimated by the PRM, conditioned on PRM hidden states. This yields structurally valid quantile estimates and enables efficient extraction of confidence bounds at arbitrary levels, which we integrate into the instance-adaptive scaling (IAS) framework of \cite{park2025know}. We evaluate on mathematical reasoning benchmarks spanning moderate-difficulty problems (MATH-500) and harder out-of-distribution problems (AIME). For PRMs with reliable ranking signals, our method substantially improves calibration over both uncalibrated PRMs and quantile regression. On downstream Best-of-N IAS performance, our method generally improves over uncalibrated PRMs. These results establish conditional optimal transport as another principled and practical approach to PRM calibration, offering structural guarantees and flexible uncertainty estimation.

Revisiting Adam for Streaming Reinforcement Learning

Florin Gogianu, Adrian Catalin Lutu, Razvan Pascanu · 2026-05-11

The authors revisit online (streaming) reinforcement learning—learning from a single pass through experience without replay buffers—focusing on how to make standard deep RL algorithms like DQN and C51 work in this setting. They find that C51 (a distributional RL algorithm) performs competitively with the recent StreamQ method, and argue that two properties matter: bounded gradients and variance-adjusted updates (like those in the Adam optimizer). Building on this, they derive an "Adaptive Q(λ)" algorithm using eligibility traces that achieves strong performance on Atari. **Main takeaways:** - Standard deep RL has relied on replay buffers for stability, but recent work (StreamQ) revisits purely online learning with eligibility traces and optimizer tweaks. - The authors find that classic algorithms (DQN, C51) with the Adam optimizer already work well in the streaming setting, without needing StreamQ's special updates. - Two key properties for robust online learning: bounded objective derivatives and variance-adjusted weight updates (Adam does the latter automatically). - Their new Adaptive Q(λ) algorithm, combining eligibility traces with variance adjustment, achieves roughly double the human baseline on 55 Atari games, outperforming prior methods.

arxiv cs.LG (Machine Learning)

arXiv:2605.06764v1 Announce Type: new Abstract: Learning from a sequence of interactions, as soon as observations are perceived and acted upon, without explicitly storing them, holds the promise of simpler, more efficient and adaptive algorithms. For over a decade, however, deep reinforcement learning walked the contrary path, augmenting agents with replay buffers or parallel sampling routines, in an effort to tame learning instability. Recently, this topic has been revisited by Elsayed et al. (2024), focusing on update computation through eligibility traces and modifications to the optimisation routine, resulting in the StreamQ algorithm. In this work we take a step back, investigating the efficacy of established updates, such as those implemented by DQN and C51 within this online setting. Not only do we find that they perform well, but through analysing how the optimisation algorithm generally, and Adam in particular, interacts with these updates, we contend that two properties are essential for robust performance: i) the derivative of the objective is to be bounded and ii) weight updates are variance-adjusted. Rigorous and exhaustive experimentation demonstrates that C51, which exhibits both characteristics, is competitive with StreamQ across a subset of 55 Atari games. Using these insights, we derive a variance-adjusted algorithm based on eligibility traces, termed Adaptive Q$(\lambda)$, which approaches double the human baseline on the same subset, surpassing existing methods by all performance metrics.

Sparse Attention as a Range Searching Problem: Towards an Inference-Efficient Index for KV Cache

Mohsen Dehghankar, Abolfazl Asudeh · 2026-05-11

The authors reframe sparse attention (selecting only a subset of key-value cache entries during LLM decoding) as a "halfspace range searching" problem from computational geometry, and build a new index called Louver that guarantees zero false negatives—every key above a threshold is retrieved—while being fast and lightweight. Louver is designed to handle the fact that the set of important tokens changes at every decoding step and for every query, which existing sparse attention methods don't guarantee. **Main takeaways:** - Sparse attention methods trade accuracy for speed, but missing even one critical key can cause sharp errors, especially in long reasoning tasks. - The authors formulate KV cache retrieval as halfspace range searching, where you want all keys whose attention score exceeds a threshold. - Louver is a new index structure that guarantees perfect recall (no false negatives) above a threshold, is easy to integrate into existing LLM pipelines, and has CPU/GPU-aware optimizations. - Experiments show Louver is more accurate than prior sparse attention methods, faster than both sparse baselines and dense FlashAttention in many settings.

arxiv cs.LG (Machine Learning)

arXiv:2605.06763v1 Announce Type: new Abstract: Sparse attention improves LLM inference efficiency by selecting a subset of key-value entries, but at the cost of potential accuracy degradation. In particular, omitting critical KV entries can induce substantial errors in model outputs. Existing methods typically operate under fixed or adaptive token budgets and provide empirical robustness or partial theoretical guarantees, yet they do not ensure zero false negatives in decoding steps, particularly since the set of relevant tokens is both query- and step-dependent. Our empirical observations confirm that missing even one critical key can lead to sharp error spikes, especially in long reasoning tasks where the set of important tokens varies throughout decoding. This observation motivates the need for indexing methods that dynamically adapt to these variations across decoding steps while guaranteeing a full recall of the relevant keys above a certain threshold. We address this challenge by reformulating sparse attention as the halfspace range searching problem. However, existing range searching indices are not suitable for modern LLM inference due to their computational and implementation overheads. To overcome this, we introduce Louver, a novel index structure tailored for efficient KV cache retrieval. Louver (i) guarantees zero false negatives with respect to a specified threshold in both theory and practice, (ii) is lightweight to integrate into existing LLM pipelines, and (iii) incorporates hardware-aware optimizations for both CPU and GPU executions. Our experiments demonstrate that Louver outperforms prior sparse attention methods in both accuracy and runtime, and is faster than highly optimized dense attentions such as FlashAttention. These results highlight that recall guarantees are a critical and overlooked dimension of sparse attention, and open a new direction for building theoretically grounded, efficient KV cache indices.

Physics-based Digital Twins for Integrated Thermal Energy Systems Using Active Learning

Umme Mahbuba Nabila, Paul Seurin, Linyu Lin, Majdi I. Radaideh · 2026-05-11

The authors build "digital twins" (real-time surrogate models) for thermal energy distribution systems by combining high-fidelity Modelica simulations with simpler surrogates (including physics-informed models like SINDyC and neural networks like GRUs) trained via active learning. Active learning intelligently selects which simulation trajectories to query, cutting the data requirement by up to 5× compared to random sampling while maintaining accuracy and enabling uncertainty quantification for real-time control. **Main takeaways:** - High-fidelity thermal system simulations are too slow for real-time control; purely data-driven surrogates can be fragile. - The authors couple system-level Modelica simulations with four surrogate types: SINDyC (sparse dynamics identification), probabilistic MvG-SINDyC, feedforward neural nets, and GRU recurrent nets. - Active learning query strategies tailored to each surrogate (e.g., Mahalanobis distance for MvG-SINDyC, prediction error for others) prioritize informative trajectories. - On a glycol heat exchanger subsystem, active learning achieves comparable accuracy with one-fifth the simulation cost; GRU is most accurate, SINDyC most efficient and interpretable, and MvG-SINDyC enables uncertainty quantification.

arxiv cs.LG (Machine Learning)

arXiv:2605.06756v1 Announce Type: new Abstract: Real-time supervisory control of thermal energy distribution systems requires digital twins that are accurate, interpretable, and uncertainty-aware, yet remain data and computationally efficient. High-fidelity simulations alone are costly, while purely data-driven surrogates often lack robustness. To address these challenges, this work proposes an active learning (AL) framework that couples system-level Modelica simulations with four simpler physics-informed and data-driven surrogate modeling approaches: deterministic Sparse Identification of Nonlinear Dynamics with Control (SINDyC), its probabilistic multivariate-Gaussian extension (MvG-SINDyC), feedforward neural network (FNN), and gated recurrent unit (GRU) network. Tailored to each surrogate, model-specific AL query strategies are employed, including Mahalanobis-distance sampling in coefficient space for MvG-SINDyC and error-based sampling in prediction space for SINDyC, FNN, and GRU, allowing the learning process to prioritize dynamically informative trajectories. The proposed approach is demonstrated on the glycol heat exchanger (GHX) subsystem of the Thermal Energy Distribution System (TEDS) at Idaho National Laboratory. Across key GHX outputs--the bypass mass flow rate $\dot{m}_{\mathrm{GHX}}$ and heat transfer rate $Q_{\mathrm{GHX}}$-the AL framework achieves comparable predictive accuracy using as few as one-fifth of the simulation trajectories required by random sampling. Among the evaluated surrogates, the GRU achieves the highest predictive fidelity, while SINDyC remains the most computationally efficient and interpretable. The probabilistic MvG-SINDyC surrogate further enables uncertainty quantification and exhibits the largest computational gains under AL.

Gradient Extrapolation-Based Policy Optimization

Ismam Nur Swapnil, Aranya Saha, Tanvir Ahmed Khan, Mohammad Ariful Haque, Ser-Nam Lim · 2026-05-11

The authors propose GXPO, a new way to update language models during reinforcement learning (like when teaching a model to reason through math problems). Standard RL updates the model using only the current gradient, while full multi-step lookahead is too expensive. GXPO approximates a longer lookahead using just three backward passes: it takes two quick optimizer steps, measures how the gradient changes, predicts where the model would land after K steps, moves partway there, and applies a correction. When the prediction becomes unstable it falls back to standard single-step updates. **Main takeaways:** - Improves accuracy on Qwen2.5 and Llama math reasoning by +1.65 to +5.00 points over standard GRPO - Achieves up to 4× step speedup and 2.33× wall-clock speedup while using only three backward passes per update - Works as a plug-in replacement for GRPO—uses the same rollouts, rewards, and advantages, no extra sampling needed - Automatically switches back to single-step updates when the extrapolation becomes unreliable

arxiv cs.LG (Machine Learning)

arXiv:2605.06755v1 Announce Type: new Abstract: Reinforcement learning is widely used to improve the reasoning ability of large language models, especially when answers can be automatically checked. Standard GRPO-style training updates the model using only the current step, while full multi-step lookahead can give a better update direction but is too expensive because it needs many backward passes. We propose Gradient Extrapolation-Based Policy Optimization (GXPO), a plug-compatible policy-update rule for GRPO-style reasoning RL. GXPO approximates a longer local lookahead using only three backward passes during an active phase. It reuses the same batch of rollouts, rewards, advantages, and GRPO loss, so it does not require new rollouts or reward computation at the lookahead points. GXPO takes two fast optimizer steps, measures how the gradients change, predicts a virtual K-step lookahead point, moves the policy partway toward that point, and then applies a corrective update using the true gradient at the new position. When the lookahead signal becomes unstable, GXPO automatically switches back to standard single-pass GRPO. We also give a plain-gradient-descent surrogate analysis that explains when the extrapolation is exact and where its local errors come from. Across Qwen2.5 and Llama math-reasoning experiments, GXPO improves the average sampled pass@1 by +1.65 to +5.00 points over GRPO and by +0.14 to +1.28 points over the strongest SFPO setting, while keeping the active-phase cost fixed at three backward passes. It also achieves up to 4.00x step speedup, 2.33x wall-clock speedup, and 1.33x backward-pass speedup in reaching GRPO's peak accuracy.

A Closed-Form Upper Bound for Admissible Learning-Rate Steps in Belief-Space Dynamics

Zixi Li, Youzhen Li · 2026-05-11

The authors derive a closed-form upper bound for how large a learning-rate step can be while still guaranteeing that an update "contracts" the model's beliefs in KL-divergence terms. Instead of treating learning rate as a hyperparameter you tune empirically, they model updates as projected steps on the probability simplex and compute the maximum step size that keeps the update well-behaved in the natural information geometry. **Main takeaways:** - Provides a formula (not a tuning heuristic) for the largest safe learning-rate step - Models updates as projected forward steps on the probability simplex - Step is "admissible" if it's contractive in KL/Bregman geometry - Positions learning-rate choice as a geometric calculation rather than pure trial-and-error

arxiv cs.LG (Machine Learning)

arXiv:2605.06741v1 Announce Type: new Abstract: Learning-rate steps are usually treated as hyperparameters. This paper isolates a local beliefspace calculation: when an update is modeled as a projected forward step on the probability simplex, admissibility means contractivity in the natural KL/Bregman geometry. Under this model, the upper bound of an admissible step is not a tuning slogan but a formula.

Geometric Kolmogorov--Arnold Network (GeoKAN)

Abhijit Sen, Bikram Keshari Parida, Giridas Maiti, Mahima Arya, Denys I. Bondar · 2026-05-11

The authors propose GeoKAN, a Kolmogorov-Arnold Network variant that learns a coordinate transformation (a diagonal Riemannian metric) to warp the input space before applying basis functions. Instead of using fixed basis functions on the raw input coordinates, GeoKAN first stretches regions with rapid variation and compresses smoother regions, reallocating representational capacity where it's most needed. They test it on physics-informed learning and differential-equation problems where functions have sharp, localized features. **Main takeaways:** - Learns a geometry-adapted coordinate system (diagonal metric) before basis expansion, rather than using fixed Euclidean coordinates - Stretches high-variation regions and compresses smooth regions, placing capacity where needed - Offers three main variants (GeoKAN-NNMetric, GeoKAN-γ, LM-KAN) and basis-specific versions (RBF, Wavelet, Fourier) - Well-suited to sharp, stiff, or strongly non-uniform regimes in scientific ML and differential equations

arxiv cs.LG (Machine Learning)

arXiv:2605.06740v1 Announce Type: new Abstract: We introduce Geometric Kolmogorov--Arnold Networks (GeoKANs), a family of geometry-aware KAN-type models in which approximation is carried out in learned, geometry-adapted coordinates rather than in fixed Euclidean input coordinates. GeoKAN achieves this by learning a diagonal Riemannian metric that warps the input before basis expansion and feature mixing. The learned metric provides a geometric inductive bias through local length scaling and volume distortion, and in physics-informed settings it also affects the differential structure seen by the model. Within this framework, we develop three main variants, namely GeoKAN-NNMetric, GeoKAN-$\gamma$, and LM-KAN. For LM-KAN, we further consider three basis-specific versions, LM-KAN-RBF, LM-KAN-Wav, and LM-KAN-Fourier. These variants allow us to study geometry-aware KAN models both as general function approximators and as surrogates in physics-informed learning. By stretching regions with rapid variation and compressing smoother regions, GeoKAN reallocates representational resolution in a task-dependent manner, allowing the model to place capacity where it is most needed. As a result, GeoKAN is well suited to sharp, stiff, localized, and strongly non-uniform regimes arising in scientific machine learning and differential-equation problems.

STDA-Net: Spectrogram-Based Domain Adaptation for cross-dataset Sleep Stage Classification

Unaza Tallal, Shruti Kshirsagar, Ankita Shukla · 2026-05-11

The authors propose STDA-Net, a framework for sleep-stage classification that works across different EEG datasets (which vary in channels, sampling rates, and recording environments). They combine a CNN on 2D spectrogram inputs, a BiLSTM to model sleep dynamics over time, and domain-adversarial training (DANN) to align source and target datasets without needing labeled target data. Tested on Sleep-EDF, SHHS-1, and SHHS-2, the approach achieves ~89% accuracy and ~88% macro F1 with lower variance than 1D baseline methods. **Main takeaways:** - Uses 2D spectrograms instead of 1D EEG signals as input to a CNN-BiLSTM architecture - Applies domain-adversarial training to align features across datasets without labeled target data - Achieves 89.03% average accuracy and 87.64% macro F1 across six cross-dataset transfer settings - Shows substantially lower variance across runs than 1D baselines, indicating better stability

arxiv cs.LG (Machine Learning)

arXiv:2605.06736v1 Announce Type: new Abstract: Accurate sleep stage classification across datasets remains challenging due to variability in EEG channel montages, sampling rates, recording environments, and subject populations. Although deep learning has shown considerable promise for automated sleep staging, most existing cross-dataset methods rely on one-dimensional EEG signal representations, whereas the use of two-dimensional spectrogram-based inputs within an unsupervised domain adaptation framework has remained largely unexplored. Here, we propose STDA-Net (Spectrogram-based Temporal Domain Adaptation Network), a framework that combines a convolutional neural network (CNN) for spectrogram-based feature extraction, a bidirectional long short-term memory (BiLSTM) module for temporal modeling of sleep dynamics, and a domain-adversarial neural network (DANN) for source-to-target feature alignment without requiring any labeled target-domain data during training. Experiments are conducted on three publicly available datasets Sleep-EDF, SHHS-1, and SHHS-2 under six cross-dataset transfer settings. Results show that the proposed framework achieves an average accuracy of 89.03% and an average macro F1-score of 87.64%, consistently outperforming existing 1D baseline methods in terms of balanced classification performance, with substantially lower variance across five independent runs, indicating improved stability and reproducibility. Overall, these findings demonstrate that 2D spectrogram-based representations, combined with temporal modeling and adversarial domain adaptation, provide a robust and competitive alternative to conventional 1D EEG inputs for cross-dataset sleep staging.

Gated QKAN-FWP: Scalable Quantum-inspired Sequence Learning

Kuo-Chung Peng, Samuel Yen-Chi Chen, Jiun-Cheng Jiang, Chen-Yu Liu, En-Jui Kuo, Yun-Yuan Wang, Prayag Tiwari, Andrea Ceschini, Chi-Sheng Chen, Yu-Chao Hsu, Chun-Hua Lin, Tai-Yue Li, Antonello Rosato, Massimo Panella, Simon See, Saif Al-Kuwari, Kuan-Cheng Chen, Nan-Yow Chen, Hsi-Sheng Goan · 2026-05-11

The authors propose gated QKAN-FWP, a recurrent architecture that uses single-qubit quantum circuits (or classical simulations of them) as learnable nonlinear activations in a Fast Weight Programmer framework. Instead of multi-qubit entanglement, they use "data re-uploading" on single qubits, which is easier to simulate classically and run on noisy quantum hardware. They test it on time-series tasks, MiniGrid RL, and solar-cycle forecasting (528-month input, 132-month forecast), where their 12.5k-parameter model beats LSTM baselines with up to 13× more parameters and runs successfully on IonQ and IBM quantum processors. **Main takeaways:** - Uses single-qubit quantum circuits as activation functions in a fast-weight recurrent framework, avoiding expensive multi-qubit gates - Introduces a scalar-gated update rule that stabilizes parameter evolution and enables parallelizable gradients - On solar-cycle forecasting, a 12.5k-parameter model outperforms LSTM baselines with 25.9k–167k parameters on MSE and peak errors - Deployed on real quantum hardware (IonQ, IBM) and recovered accuracy within 0.1% relative MSE at 1024 shots

arxiv cs.LG (Machine Learning)

arXiv:2605.06734v1 Announce Type: new Abstract: Fast Weight Programmers (FWPs) encode temporal dependencies through dynamically updated parameters rather than recurrent hidden states. Quantum FWPs (QFWPs) extend this idea with variational quantum circuits (VQCs), but existing implementations rely on multi-qubit architectures that are difficult to scale on noisy intermediate-scale quantum (NISQ) devices and expensive to simulate classically. We propose gated QKAN-FWP, a fast-weight framework that integrates FWP with Quantum-inspired Kolmogorov-Arnold Network (QKAN) using single-qubit data re-uploading circuits as learnable nonlinear activation, known as DatA Re-Uploading ActivatioN (DARUAN). We further introduce a scalar-gated fast-weight update rule that stabilizes parameter evolution, supported by a theoretical analysis of its adaptive memory kernel, geometric boundedness, and parallelizable gradient paths. We evaluate the framework across time-series benchmarks, MiniGrid reinforcement learning, and highlight real-world solar cycle forecasting as our main practical result. In the long-horizon setting with 528-month input window and 132-month forecast horizon, our 12.5k-parameter model achieves lower scaled Mean Square Error (MSE), peak amplitude error, and peak timing error than a suite of classical recurrent baselines with up to 13x more parameters, including Long Short-Term Memory (LSTM) networks (25.9k-89.1k parameters), WaveNet-LSTM (167k), Vanilla recurrent neural network (11.5k), and a Modified Echo State Network (132k). To validate NISQ compatibility, we further deploy the trained fast programmer on IonQ and IBM Quantum processors, recovering forecasting accuracy within 0.1% relative MSE of the noiseless simulator at 1024 shots. These results position gated QKAN-FWP as a scalable, parameter-efficient, and NISQ-compatible approach to quantum-inspired sequence modeling.

Beyond Factor Aggregation: Gauge-Aware Low-Rank Server Representations for Federated LoRA

Jinqian Chen, Chang Liu, Jihua Zhu · 2026-05-11

The authors propose GLoRA, a federated-learning approach for LoRA fine-tuning that fixes a conceptual problem: directly averaging LoRA factors (the A and B matrices) is "gauge-dependent"—you can rotate or rescale them arbitrarily without changing the actual update, so factor averaging can produce different results depending on the coordinate system. GLoRA instead estimates a consensus update subspace from clients' projectors and aggregates updates in shared coordinates, making aggregation semantically meaningful. It also supports heterogeneous client ranks (different clients can use different LoRA ranks) by projecting onto a shared server representation. **Main takeaways:** - Standard federated LoRA averages factors directly, but this is coordinate-dependent—the same update can be written infinitely many ways - GLoRA estimates a consensus subspace and aggregates updates in shared reference coordinates, making aggregation gauge-invariant - Supports heterogeneous client ranks: clients can use different LoRA ranks and GLoRA projects them onto a shared server state - Outperforms federated LoRA baselines on GLUE and SuperNI under data, resource, and task heterogeneity

arxiv cs.LG (Machine Learning)

arXiv:2605.06733v1 Announce Type: new Abstract: Federated LoRA enables parameter-efficient adaptation of large language models under decentralized data and limited client resources.However, directly averaging LoRA factors is representation-dependent: the same intrinsic update admits infinitely many gauge-equivalent factorizations, so factor-level aggregation can change under arbitrary coordinate choices while the underlying update remains unchanged. This reveals a semantic mismatch in existing federated LoRA aggregation rules. We propose \textbf{GLoRA}, a gauge-aware server representation for federated LoRA.Instead of aggregating raw factors, GLoRA estimates a consensus update subspace from client projectors and aggregates client updates in shared reference coordinates, thereby representing semantic update aggregation entirely in low-rank form. To support heterogeneous client capacities, GLoRA further provides a rank-compatible readout that instantiates adapters of different ranks from the same server state without dense update reconstruction. Experiments on GLUE and SuperNI show that GLoRA consistently outperforms federated LoRA baselines under data, resource, and task heterogeneity, including heterogeneous client ranks, sparse participation, larger backbones, and unseen-task evaluation. GLoRA also achieves a favorable efficiency--performance trade-off, suggesting that effective federated LoRA requires not merely averaging low-rank factors, but defining a semantically meaningful server-side representation for aggregation.

On Training in Imagination

Nadav Timor, Ravid Shwartz-Ziv, Micah Goldblum, Yann LeCun, David Harel · 2026-05-11

The authors study model-based RL, where policies are trained on imagined rollouts from learned dynamics and reward models without querying the real environment. They analyze how errors in the learned models affect returns and derive the optimal sample allocation—how many samples to spend learning dynamics vs. rewards—under power-law scaling. They also show that zero-mean reward noise doesn't bias the REINFORCE gradient, only adds variance, which creates a tradeoff: is it better to get more rollouts with cheap noisy rewards, or fewer rollouts with expensive accurate rewards? **Main takeaways:** - Extends error-propagation analysis to MDPs with learned reward models and derives optimal dynamics-vs-reward sample allocation - Identifies lower Lipschitz constants (smoother dynamics, reward, and policy) as a representation goal that tightens error bounds - Shows zero-mean reward noise leaves REINFORCE gradient unbiased, only adding variance that decreases with more rollouts - Frames the noisy-reward tradeoff as a one-dimensional optimization: more cheap noisy rollouts vs. fewer expensive accurate ones

arxiv cs.LG (Machine Learning)

arXiv:2605.06732v1 Announce Type: new Abstract: State-of-the-art model-based reinforcement learning methods train policies on imagined rollouts. These rollouts are trajectories generated by a learned dynamics model and are scored by a learned reward model, but without querying the true environment during policy updates. We study this training paradigm by quantifying how errors in learned dynamics and reward models affect returns and policy optimization. First, we extend the analysis of Asadi et al. (2018) to MDPs with learned reward models, and derive the optimal sample allocation--the ratio of dynamics samples to reward samples that minimizes a bound on return error under power-law scaling assumptions. We identify lower Lipschitz constants of the learned dynamics, reward, and policy as a representation desideratum that tightens this bound, and we connect this perspective to the temporal-straightening objective of Wang et al. (2026). Second, we examine how policy optimization with REINFORCE tolerates noisy rewards, which are often cheaper to obtain. We show that zero-mean reward noise leaves the gradient estimator unbiased and adds at most a variance term that decreases with the number of rollouts. This introduces a practical tradeoff: given a fixed budget, should one buy more rollouts with cheaper but noisier rewards, or fewer rollouts with more expensive but less noisy rewards? We reduce this choice to a one-dimensional optimization problem and characterize the optimum.

Semantic State Abstraction Interfaces for LLM-Augmented Portfolio Decisions: Multi-Axis News Decomposition and RL Diagnostics

Likhita Yerra (AIVANCITY School of AI and Data), Remi Uttejitha Allam (AIVANCITY School of AI and Data) · 2026-05-11

The authors introduce SSAI, a framework for mapping sparse unstructured text (news) into K named, auditable coordinates (e.g., sentiment, risk, confidence, volatility) with neutral defaults on no-news days. They use it to build trading portfolios on US equities, testing direct factor portfolios, supervised ridge forecasters, and RL agents that all share the same four-axis representation. The four-factor portfolio reaches 307% cumulative return and Sharpe 1.067, but the apparent edge over buy-and-hold (243%) fails when you control for coverage stratification, apply realistic transaction costs (≥0.2%), or compare to simpler baselines like sentiment-only or PCA. **Main takeaways:** - Proposes a framework for converting sparse text into K auditable dimensions with neutral defaults, separating representation design from optimizer variance - Tests on 30 NASDAQ-100 stocks (2019–2023) with four axes: sentiment, risk, confidence, volatility forecast - Four-factor portfolio shows 307% return and Sharpe 1.067, but edge over buy-and-hold is fragile under controls and transaction costs - Sentiment-only baseline, PC1 composite, and FinBERT portfolio are stronger ranking signals in this setting

arxiv cs.LG (Machine Learning)

arXiv:2605.06730v1 Announce Type: new Abstract: We introduce Semantic State Abstraction Interfaces (SSAI): a methodological template for mapping sparse unstructured text into $K$ auditable, named coordinates with neutral defaults on no-news days, designed to separate representation hypotheses from optimisation variance in sequential decision systems. Our contribution is the framework and its evaluation protocol, not a claim that SSAI outperforms denser alternatives. We instantiate SSAI with $K=4$ axes (sentiment, risk, confidence, volatility forecast) on a US-equity panel (30 NASDAQ-100 names, FNSPID news, 2019--2023 test), and evaluate it across direct factor portfolios, supervised ridge forecasters, and RL agents (DP-PPO, SAC) that share the same fixed $\phi$. The four-factor factor portfolio reaches 307.2% cumulative return and Sharpe 1.067, but apparent gains versus buy-and-hold (243.6%) fail coverage-stratified controls, reverse at $\geq 0.2$% costs, and are statistically fragile versus a sentiment-only baseline; a PC1 composite and a FinBERT portfolio baseline are stronger ranking signals in this setting. Ridge and RL blocks diagnose representation versus optimiser effects. We position SSAI as an interpretability-performance diagnostic and reusable protocol for sparse-text decision systems.

The E$Δ$-MHC-Geo Transformer: Adaptive Geodesic Operations with Guaranteed Orthogonality

Arash Shahmansoori · 2026-05-11

The authors build a Transformer variant that keeps intermediate representations on a curved geometry (a manifold) using rotation and reflection operations that preserve distances and angles. The key innovation is a way to combine two different mathematical tricks (Cayley transforms for rotation, Householder reflections for flipping) with a learned gate that decides which operation to use, ensuring the network never accidentally distorts its representations. They show this hybrid approach gives better stability during training and handles both types of orthogonal transformations (rotations and reflections) where previous methods could only do one or the other. **Main takeaways:** - Orthogonal operations (transformations that preserve vector lengths and angles) help neural networks stay stable during training, especially over long sequences - Standard methods like Deep Delta Learning only guarantee orthogonality at specific settings; this work maintains it for all parameter values by using Cayley transforms - A learned gate decides whether to rotate or reflect representations at each layer, with a penalty encouraging clean either/or choices rather than blending - In experiments with ~1.8M parameters, the architecture showed 1.9× better long-term stability than a competing method and 3.8× better than GPT-style baselines - The method needs 33% fewer layers to achieve similar performance to baselines

arxiv cs.LG (Machine Learning)

arXiv:2605.06729v1 Announce Type: new Abstract: We present the E$\Delta$-MHC-Geo Transformer, a novel architecture that unifies Manifold-Constrained Hyper-Connections (mHC), Deep Delta Learning (DDL), and the Cayley transform to obtain input-adaptive, unconditionally orthogonal residual connections. Unlike DDL, whose Householder operator is orthogonal only at $\beta \in \{0,2\}$, our Data-Dependent Cayley rotation $Q(x)=(I+(\beta/2)A(x))^{-1}(I-(\beta/2)A(x))$ preserves orthogonality for all $\beta$ and all inputs. To handle negation, an eigenvalue $-1$ case that Cayley provably excludes, we introduce the E$\Delta$-MHC-Geo Hybrid, which combines Cayley rotation with Householder reflection via a learned operator-selection gate $X'=\gamma(X)Q(X)X+(1-\gamma(X))H_2(X)X$. A midpoint-collapse regularizer, $4\gamma(1-\gamma)$, encourages boundary gate decisions, where each selected component is orthogonal. In matched-parameter comparisons, with approximately 1.79M parameters per model and mean +/- standard deviation over 3 seeds, against four baselines including the concurrent JPmHC, E$\Delta$-MHC-Geo achieves the best long-horizon stability, 1.9x over JPmHC and 3.8x over GPT; the best near-$\pi$ rotation loss, 4.5x over JPmHC on single-plane; strong norm preservation, with 0.001 mean deviation; and 0.96 negation cosine alignment in a diagnostic reflection probe, all with 33% fewer layers. While JPmHC's wider representation excels on pure rotation, its finite Cayley residual mixer excludes an exact $\lambda=-1$ operator and has no reflection branch, motivating our hybrid approach for accessing both connected components of $O(n)$.

Medical Imaging Classification with Cold-Atom Reservoir Computing using Auto-Encoders and Surrogate-Driven Training

Nuno Batista, Ana Morgado, Oscar Ferraz, Sagar Silva Pratapsi, Jorge Lobo, Gabriel Falcao · 2026-05-11

The authors built a medical imaging classifier (for detecting polyps in images) that uses a quantum computing simulator as one of its processing layers. Because quantum measurements aren't differentiable (you can't use standard backpropagation through them), they train a surrogate neural network that mimics the quantum layer, allowing gradients to flow end-to-end. An autoencoder compresses the high-dimensional images into compact representations that are fed into the quantum simulator, and the whole pipeline is trained jointly for both accurate classification and faithful image reconstruction. **Main takeaways:** - Quantum computing layers can't be trained with standard backpropagation because measurements are discrete and non-differentiable - A differentiable surrogate model (a neural net trained to emulate the quantum layer) bridges the gradient barrier and enables end-to-end training - An autoencoder reduces image dimensionality before the quantum step; it's trained jointly for both compression and classification accuracy - The quantum embeddings come from expectation values in a simulated Rydberg Hamiltonian (a physics model of interacting atoms) - The hybrid pipeline outperforms traditional PCA or unguided autoencoders on polyp detection, even with current noisy quantum simulators

arxiv cs.LG (Machine Learning)

arXiv:2605.06727v1 Announce Type: new Abstract: We introduce a hybrid quantum-classical pipeline, based on neutral-atom reservoir computing, for medical image classification, focusing on the binary classification task of polyp detection. To deal effectively with the high dimensionality, we integrate a guided auto-encoder. This pipeline learns compact and discriminative representations of image data that are also well-suited for quantum reservoir computing. A key challenge in such systems is the non-differentiable nature of quantum measurements, which creates a 'gradient barrier' for standard training. We overcome this barrier by incorporating a differentiable surrogate model that emulates the quantum layer, enabling end-to-end backpropagation through the entire system. This guided training process is jointly optimized for classification accuracy and for faithful image recovery from the auto-encoder. The learned latent representations are encoded as pulse detuning parameters within a Rydberg Hamiltonian, and quantum embeddings are subsequently obtained through expectation values. These embeddings are then passed to a linear classifier. Our simulations show that this method outperforms some traditional approaches that use PCA or unguided autoencoders. We also conduct ablation studies to assess the impact of various quantum and training parameters, demonstrating the robustness and flexibility of our proposed pipeline for real-world medical imaging applications, even in the current NISQ era.

Transformer-Based Wildlife Species Classification from Daily Movement Trajectories

Obed Irakoze, Prasenjit Mitra · 2026-05-11

The authors train sequence models to identify animal species from GPS movement trajectories alone, using data from seven species tracked on Movebank. They test how well models generalize when entire studies or geographic regions are held out (a harder test than random splitting). Transformers beat LSTM, CNN, and temporal convolutional baselines by 8–22 percentage points in balanced accuracy, and adding richer movement features (speed, direction, turning angle) beyond simple displacement helps especially for rare species like lions and zebras. **Main takeaways:** - Transformers consistently outperform recurrent and convolutional architectures on wildlife trajectory classification, achieving 0.83 balanced accuracy and 0.92 AUC on elephant detection - Testing with whole studies or regions held out is much harder than random splits, but better reflects real deployment scenarios - Enriching trajectories with speed, direction, and turning features (not just position changes) significantly improves performance, especially for underrepresented species - One-hour resolution works better than 30-minute resolution across studies, likely because it reduces missing data and ensures consistent temporal coverage - The approach shows promise for species identification from movement patterns alone, useful when visual identification is impractical

arxiv cs.LG (Machine Learning)

arXiv:2605.06726v1 Announce Type: new Abstract: Inferring the identity of wildlife species from daily movement data alone is a challenging task. We train sequence models on large-scale, 7-species GPS trajectories from the Movebank platform. Trajectories models are evaluated using a protocol in which entire telemetry studies or regions are heldout during testing. We compare Transformer-based sequence models to LSTM, CNN, and Temporal Convolutional Networks, and find that Transformers consistently achieve higher balanced accuracy with gains of approximately 8 to 22 percentage points, depending on the species and experimental setting. In an elephant binary classification task with 1-hour resolution, the Transformer achieves a balanced accuracy of 0.83 and an AUC of 0.92, substantially outperforming all baseline models. We examine, under data-limited conditions, feature representations by analyzing the differences between a basic displacement-based encoding and an expanded range of movement descriptors that include speed, direction, and turning behavior. With feature augmentation, we see clear performance gains, especially for underrepresented and sparsely represented species, such as large carnivores, lions, and Zebras. Finally, experiments comparing 1-hour and 30-minutetemporal resolutions show that while finer sampling can capture short-term movement patterns for some species, a unified 1-hour resolution yields more promising performance across studies by reducing missing data and ensuring consistent temporal coverage.

Enabling Unsupervised Training of Deep EEG Denoisers With Intelligent Partitioning

Qiyu Rao, Haozhe Tian, Homayoun Hamedmoghadam, Danilo Mandic · 2026-05-11

The authors develop a self-supervised method for training deep learning models to denoise EEG signals from wearable sensors, which is hard because you never have truly clean EEG to use as training targets. Their key idea (Intelligent Partitioning for Self-supervised Denoising, iPSD) is to learn how to split a single noisy EEG segment into independent chunks that share the same underlying brain signal but have different noise realizations, then train the denoiser to produce consistent outputs across these partitions. This works even in zero-shot settings where you only have one segment to denoise. **Main takeaways:** - Traditional EEG denoising methods use fixed rules that can't handle time-varying artifacts; deep learning methods need clean training data that doesn't exist - iPSD eliminates the need for clean reference signals by partitioning a single noisy input into independent realizations with the same underlying signal - The method enables self-supervised training where the denoiser learns to produce outputs consistent across different noisy views of the same signal - Works in extremely challenging conditions: signal-to-noise ratios down to -10 dB (signal 10× weaker than noise) and muscle artifacts - Achieves state-of-the-art performance on wearable in-ear EEG with spectral fidelity orders of magnitude better than baselines

arxiv cs.LG (Machine Learning)

arXiv:2605.06724v1 Announce Type: new Abstract: Denoising wearable electroencephalogram (EEG) is inherently challenging since neural activity is not only subtle but also inseparable from spectrally overlapping noise artifacts. Classical signal processing methods, relying on fixed or heuristic rules, cannot handle the time-varying pervasive artifacts in wearable EEGs. Deep learning methods, on the other hand, show promise in decomposition-free EEG denoising using highly expressive neural networks, but the training requires artifact-free EEG, which is inherently unobtainable. To address this, we propose Intelligent Partitioning for Self-supervised Denoising (iPSD). Our method eliminates the need for clean references by learning to partition an input EEG segment into independent noisy realizations with the same underlying signal. This enables self-supervision of deep learning denoisers, even in zero-shot settings where only a single EEG segment to be denoised is available. We validate iPSD through extensive experiments, including validations on wearable EEG from in-ear sensors. The results show that iPSD achieves state-of-the-art performance, most notably under extremely low signal-to-noise ratios (down to -10 dB) and challenging artifacts (e.g., EMG), with spectral fidelity orders of magnitude higher than competitive baselines.

Conditional generation of antibody sequences with classifier-guided germline-absorbing discrete diffusion

Justin Sanders, Luca Giancardo, Lan Guo, Yue Zhao, Kemal Sonmez, Nina Cheng, Melih Yilmaz · 2026-05-11

The authors tackle antibody design using a modified diffusion model that avoids simply memorizing germline (inherited template) sequences. Standard protein language models trained on antibody data tend to just reproduce germline sequences rather than learning the biologically meaningful mutations that happen during immune response. Their "germline-absorbing diffusion" uses the germline sequence as the starting/ending point of the diffusion process (instead of random noise), forcing the model to learn only the trajectory of somatic mutations. This dramatically improves the model's ability to predict non-germline mutations (from 26% to 46% accuracy) and enables better conditional generation of antibodies with desired properties like reduced hydrophobicity. **Main takeaways:** - Standard protein language models memorize germline (inherited) antibody sequences rather than learning the somatic mutations that create useful antibodies - Germline-absorbing diffusion modifies the noise process so the model learns the trajectory from germline to mutated sequence, excluding genetic recombination statistics - This biological inductive bias improves non-germline residue prediction from 26% to 46%, approaching the theoretical limit set by natural biological variability - The model supports classifier-guided generation: you can condition antibody design on any property predicted by an off-the-shelf classifier - Demonstrates improved trade-offs between target properties (hydrophobicity, binding affinity) and sample quality compared to baselines

arxiv cs.LG (Machine Learning)

arXiv:2605.06720v1 Announce Type: new Abstract: Antibody therapeutics are among the most successful modern medicines, yet computationally designing antibodies with desirable binding and developability properties remains challenging. While protein language models (pLMs) have emerged as powerful tools for antibody sequence design, existing approaches largely suffer from two key limitations: they predominantly memorize germline sequences rather than modeling biologically meaningful somatic variation, and they offer limited support for flexible classifier-guided conditional generation. We address these challenges through two primary contributions. First, we demonstrate that discrete diffusion fine-tuning achieves strong language modeling performance on antibody sequences while allowing for generation conditioned on any off-the-shelf classifier. Second, we introduce germline absorbing diffusion, a novel modification of the discrete diffusion noise process in which the germline sequence - rather than a masked sequence - serves as the absorbing state. This biologically motivated inductive bias restricts the model to learning the trajectory from germline to observed sequence, effectively excluding genetic variation and V(D)J recombination statistics from the learned distribution and dramatically mitigating germline bias. We show that germline diffusion improves non-germline residue prediction accuracy from 26 percent to 46 percent, approaching the theoretical upper bound set by true biological variability. We then demonstrate the utility of our germline diffusion model on the conditional generation tasks of sampling antibodies with improved hydrophobicity and predicted binding affinity. On both tasks our model shows an improved tradeoff between class adherence and sample quality, significantly outperforming EvoProtGrad, a popular strategy to sample from pLMs with gradient-based discrete Markov Chain Monte Carlo.

Robustness of Refugee-Matching Gains to Off-Policy Evaluation Choices

Kirk Bansak, Elisabeth Paulson, Dominik Rothenh\"ausler, Jeremy Ferwerda, Jens Hainmueller, Michael Hotard · 2026-05-11

The authors test whether earlier findings about algorithmic refugee matching (which suggested better employment outcomes when refugees are matched to locations using predicted outcomes) hold up when you change the evaluation method. They re-analyze the same US refugee data using several different off-policy evaluation techniques—ways to estimate what would have happened under a different assignment policy without actually running that policy. The core result is that the estimated gains from algorithmic matching stay consistent across methods: all approaches show similar positive impacts, and most reach statistical significance. **Main takeaways:** - Off-policy evaluation estimates the impact of a counterfactual policy (algorithmic refugee matching) using data from the actual policy (administrative assignments) - Multiple estimation methods (inverse probability weighting, augmented inverse probability weighting variants) all produce similar impact estimates - The robustness holds across different modeling architectures and assignment procedures - Results remain consistent with the original 2018 Bansak et al. findings, strengthening confidence in the refugee-matching gains - Most estimation approaches achieve statistical significance, though the exact confidence varies by method

arxiv cs.LG (Machine Learning)

arXiv:2605.06686v1 Announce Type: new Abstract: Previous research has investigated the potential of refugee matching for boosting refugee outcomes, first considered by Bansak et al. (2018). This paper demonstrates the stability of counterfactual impact evaluation results in the context of refugee matching in the United States using a range of off-policy evaluation methods. In order to estimate counterfactual impact and test the robustness of our results, we employ several evaluation methods, including inverse probability weighting (IPW) and multiple variants of augmented inverse probability weighting (AIPW). We also consider various modifications, including alternative modeling architectures and different assignment procedures. The impact estimates remain consistent in magnitude in all scenarios as well as statistically significant in most cases. Furthermore, the estimates are also consistent with the results originally presented in Bansak et al. (2018).

From Canopy to Collision: A Hybrid Predictive Framework for Identifying Risk Factors in Tree-Involved Traffic Crashes

Abdul Azim, Ahmed Hossain, Soumyadip Maitra, Panick Kalambay · 2026-05-11

The authors analyze traffic crashes involving trees (a subset of run-off-road collisions that tend to be especially deadly) to identify what factors make them more severe. They use a gradient-boosting classifier (CatBoost) to predict crash severity, SHAP values to explain which factors matter most, logistic regression to validate the findings, and SHAP interaction plots to find combined effects. Not wearing a seatbelt is the strongest predictor—unrestrained occupants are nearly 3× more likely to have severe outcomes—followed by vehicle age, speeding, and driver impairment. **Main takeaways:** - Restraint non-use is the dominant risk factor: unrestrained occupants are nearly three times more likely to experience fatal or incapacitating injuries, largely due to ejection risk - Vehicle age, speeding violations, and driver impairment all substantially increase severity through reduced crashworthiness, higher impact forces, and impaired control - Key risk interactions emerge: poor lighting with older vehicles, speeding with poor lighting, no restraints with older vehicles, and wet roads with speeding all show additive effects - SHAP values (a model-agnostic explanation method) and logistic regression coefficients largely agree on factor importance, cross-validating the findings - Results support targeted interventions like enhanced seat belt enforcement, speed management, and roadside hazard mitigation

arxiv cs.LG (Machine Learning)

arXiv:2605.06684v1 Announce Type: new Abstract: Tree-involved crashes represent a critical subset of run-off-road (ROR) collisions, often resulting in fatal or severe injuries due to high-energy impacts. This study develops a comprehensive analytical framework to identify and quantify risk factors contributing to crash severity in tree-involved collisions using the Crash Report Sampling System (CRSS) database spanning 2020-2023. The modeling framework follows a multi-step process. First, a machine learning based classification model (CatBoost) identifies key factors associated with binary crash injury severity (KA: fatal or incapacitating injury versus BC: non-incapacitating or possible injury). Second, SHapley Additive exPlanations (SHAP) tool is used to quantify and visualize the marginal effects of top influential factors on crash severity. Third, a binary logistic regression model estimates factor effects and validates SHAP-derived importance measures. Finally, SHAP interaction plots examine the combined effects of key contributing factors. Results reveal restraint non-use as the most influential predictor, with unrestrained occupants nearly three times more likely to experience severe outcomes due to ejection risk. Vehicle age, speeding violations, and driver impairment demonstrate substantial effects, reflecting reduced crashworthiness, increased impact forces, and reduced control capabilities. Critical interactions emerge between lighting conditions and vehicle age, speeding and lighting conditions, restraint use and vehicle age, and road surface and speeding, demonstrating additive risk effects with specific interactions. These findings provide critical insights for targeted safe system-based interventions, including enhanced seat belt enforcement, speed management in reduced visibility conditions, and vehicle fleet modernization.

Toeplitz MLP Mixers are Low Complexity, Information-Rich Sequence Models

Benjamin L. Badger, Ethan Roland · 2026-05-11

The authors propose Toeplitz MLP Mixers (TMMs), a Transformer alternative that replaces attention with multiplication by Toeplitz matrices (structured matrices where diagonals are constant). This drops computational complexity from quadratic to O(n log n) during training and linear at inference, while using less memory. Despite lacking the sophisticated gating and state tracking of architectures like Mamba, TMMs retain more information from inputs, showing better copying ability and in-context learning. The authors argue this comes from fewer architectural biases, and they find that trained Toeplitz layers tend to be nearly invertible even though the overall model isn't required to be. **Main takeaways:** - Toeplitz matrix multiplication (where each diagonal has the same value) replaces quadratic-cost attention with O(n log n) training and O(n) inference complexity - TMMs lack explicit gating or state maintenance but achieve better training efficiency (lower loss per compute and memory usage) than comparable architectures - Higher information retention from inputs leads to improved copying, retrieval, and in-context learning compared to other sub-quadratic alternatives - Counterintuitively, trained Toeplitz layers are more likely to be invertible (or nearly so) than in architectures designed to be invertible - The authors argue fewer architectural biases allow the model to retain more input information, benefiting tasks requiring faithful recall

arxiv cs.LG (Machine Learning)

arXiv:2605.06683v1 Announce Type: new Abstract: Transformer-based large language models are in some respects limited by the quadratic time and space computational complexity of attention. We introduce the Toeplitz MLP Mixer (TMM), a transformer-like architecture that swaps attention for triangular-masked Toeplitz matrix multiplication over the sequence dimension resulting in $\mathcal{O} (dn \log n)$ time and $\mathcal O(dn)$ space complexity during training and $\mathcal O(dn)$ time and space at inference prefill. Despite the lack of sophisticated input modulation or state maintenance present in other sub-quadratic architectures, TMMs yield greater training efficiency in terms of loss achieved per compute and device memory. We demonstrate that TMMs are capable of retaining more input information resulting in improved copying ability, which we argue results from a lack of architectural biases. Consistent with higher input information retention, TMMs exhibit superior information retrieval and in-context learning benchmark accuracy compared to comparable architectures. We conclude with an analysis from the perspective of operator index theory and show that, counterintuitively, trained Toeplitz layers of causal non-invertible models are more likely to be invertible or nearly so than models that are actually invertible over their inputs.

A Hierarchical Ensemble Pipeline for Anomaly Detection in ESA Satellite Telemetry

Lorenzo Riccardo Allegrini, Geremia Pompei · 2026-05-11

The authors built a multi-stage anomaly detection system for satellite telemetry from the European Space Agency. Their pipeline extracts features from each sensor channel separately, stacks models within channels, then aggregates predictions across channels. They use careful cross-validation with masking to prevent the model from "peeking" at future data, and show strong results on ESA's anomaly detection benchmark. **Main takeaways:** - Hierarchical approach: model each sensor independently first, then combine predictions across sensors - Two-level masking during training prevents information leakage from future timesteps - Combines shapelet features (pattern-based) with statistical features - Strong performance on realistic satellite data with subtle anomalies - Time-series cross-validation ensures the system generalizes to new periods

arxiv cs.LG (Machine Learning)

arXiv:2605.06681v1 Announce Type: new Abstract: A hierarchical ensemble pipeline is introduced to address anomaly detection in multivariate telemetry data provided by European Space Agency (ESA). The method integrates shapelet-based and statistical feature extraction, per-channel modeling, intra-channel stacking, and a final cross-channel aggregation. The pipeline is trained and validated using time-series cross-validation and two-level masking strategies to prevent information leakage. Results on the European Space Agency Anomaly Detection Benchmark (ESA-ADB) challenge demonstrate strong generalization, highlighting the effectiveness of hierarchical modeling in detecting subtle anomalies in realistic satellite telemetry.

On the Role of Strain and Vorticity in Numerical Integration Error for Flow Matching

Chenxi Tao, Seung-Kyum Choi · 2026-05-11

The authors analyze why flow matching models (which generate data by integrating a learned velocity field) need many integration steps. They split the velocity field's derivative into two parts: strain (how the flow stretches) and vorticity (how it rotates). Strain causes errors to grow exponentially during integration, while vorticity only adds linear error. They show optimal transport flows have zero vorticity and propose regularizing strain more heavily than vorticity, cutting integration error by 2.7× on synthetic data. **Main takeaways:** - Flow matching integration error comes mainly from strain (stretching), not vorticity (rotation) - Optimal transport velocity fields are rotation-free and can be integrated exactly with simple Euler steps - Weighted regularization that penalizes strain more than vorticity reduces the number of steps needed - On CIFAR-10, lightweight fine-tuning with this regularization improves image quality by 14% at 10 integration steps - The math is grounded in the Jacobian decomposition of the velocity field

arxiv cs.LG (Machine Learning)

arXiv:2605.06680v1 Announce Type: new Abstract: Flow matching generates data by integrating a learned velocity field, where the number of integration steps (NFE) directly determines inference cost. We analyze which properties of the velocity field govern integration error by decomposing the velocity Jacobian into its symmetric part S (strain rate) and antisymmetric part Omega (vorticity). We prove that strain and vorticity play different roles: strain controls exponential error amplification through the logarithmic norm, while vorticity contributes only linearly to the local truncation error. We further show that the optimal transport velocity field is irrotational and has zero material derivative, implying second-order Euler accuracy; for exact displacement interpolation, the associated Lagrangian particle dynamics are integrated exactly by Euler. Motivated by this analysis, we study weighted Jacobian regularization with strain weight alpha and vorticity weight beta. Experiments on 2D synthetic data confirm the main theoretical predictions, showing up to 2.7x lower integration error at NFE=5. Preliminary CIFAR-10 experiments show consistent trends, with a lightweight fine-tuning procedure improving FID by 14 percent at NFE=10 while preserving high-NFE quality.

Breaking the Illusion: When Positive Meets Negative in Multimodal Decoding

Yubo Jiang, Yitong An, Xin Yang, Abudukelimu Wuerkaixi, Xuxin Cheng, Fengying Xie, Zhiguo Jiang, Cao Liu, Ke Zeng, Haopeng Zhang · 2026-05-11

Vision-language models often "hallucinate" objects that aren't in the image because they rely too heavily on language patterns and under-weight visual information. The authors introduce Positive-and-Negative Decoding (PND), a training-free method that intervenes during text generation. It runs two parallel paths: one that amplifies visual signals and one that constructs counterfactuals to suppress language-prior-driven outputs. By contrasting these paths at each token, PND forces the model to stick to what's actually in the image, achieving state-of-the-art hallucination reduction. **Main takeaways:** - Vision-language models under-weight visual features relative to language priors during generation - PND is training-free — it only changes how tokens are selected during inference - Dual paths: positive path boosts visual evidence, negative path penalizes prior-driven guesses - Achieves best results on POPE, MME, and CHAIR hallucination benchmarks - The core insight is attention imbalance: visual attention scores are too low

arxiv cs.LG (Machine Learning)

arXiv:2605.06679v1 Announce Type: new Abstract: Vision-Language Models (VLMs) are frequently undermined by object hallucination, generating content that contradicts visual reality, due to an over-reliance on linguistic priors. We introduce Positive-and-Negative Decoding (PND), a training-free inference framework that intervenes directly in the decoding process to enforce visual fidelity. PND is motivated by our finding of an attention imbalance in VLMs, where visual features are under-weighted. Our framework introduces a dual-path contrast: a positive path that amplifies visual evidence and a negative path that constructs counterfactuals to penalize prior-dominant generation. By contrasting outputs from both paths during decoding, PND steers generation toward visually grounded results. Experiments on POPE, MME, and CHAIR demonstrate state-of-the-art performance without retraining.

A Wasserstein GAN-based climate scenario generator for risk management and insurance: the case of soil subsidence

Antoine Heranval (BioSP), Olivier Lopez (CREST), Didier Ngatcha (CREST), Daniel Nkameni (CREST) · 2026-05-11

The authors use conditional GANs to generate future scenarios of drought conditions (measured by a Soil Wetness Index) in France up to 2050. Their model, SwiGAN, produces realistic spatial maps of how drought evolves over time, helping insurance companies plan for climate-driven risks beyond the usual one-year horizon. Drought accounts for 30% of natural catastrophe insurance payouts in France, making long-term forecasting valuable. **Main takeaways:** - Conditional GANs generate plausible future spatial maps of soil wetness under climate change - Focused on France, where drought is a major insurance expense - Helps insurers design strategies for climate risk beyond short-term regulations - The approach is generalizable to other climate perils and actuarial problems - Trained on historical soil wetness data to learn realistic drought propagation patterns

arxiv cs.LG (Machine Learning)

arXiv:2605.06678v1 Announce Type: new Abstract: According to the United Nations Office for Disaster Risk Reduction (2025), the average annual cost of natural catastrophes increased from 70--80 billion USD between 1970 and 2000 to 180--200 billion USD between 2001 and 2020. Reports from organizations such as the IFOA and the WWF highlight the need for the insurance sector to adapt to this rapidly evolving context by developing medium- to long-term strategies that go beyond the one-year horizon of prudential regulations such as Solvency II. This paper introduces an artificial intelligence framework based on Conditional Generative Adversarial Networks (Conditional GANs) to generate future spatio-temporal trajectories of climatic indices. The approach focuses on the Soil Wetness Index (SWI), a key indicator used in France to assess drought severity. Drought accounts for approximately 30% of the indemnities paid under the French natural catastrophe insurance scheme. The proposed model, SwiGAN, simulates plausible drought propagation patterns up to 2050 for a region of France particularly exposed to this hazard. By generating realistic sequences of SWI maps, SwiGAN provides insights into drought dynamics under climate change scenarios and supports the design of adaptive risk management and insurance strategies. The methodology is also generalizable to other climate-related perils and actuarial applications such as economic scenario generation.

LKV: End-to-End Learning of Head-wise Budgets and Token Selection for LLM KV Cache Eviction

Enshuai Zhou, Yifan Hao, Chao Wang, Rui Zhang, Di Huang, Jiaming Guo, Xing Hu, Zidong Du, Qi Guo, Yunji Chen · 2026-05-11

Long-context language models store key-value pairs for every past token, and this cache becomes a memory bottleneck. Existing compression methods use hand-crafted rules to decide which tokens to keep (like "attention sinks" or recent tokens), but these rules don't optimize for the actual task. LKV learns both which attention heads need more cache budget and which tokens matter most, treating compression as an end-to-end optimization problem. On LongBench, it retains only 15% of the cache with almost no quality loss, and the learned budget allocation turns out to be more important than the token-selection method. **Main takeaways:** - Existing KV cache compression uses heuristics (recent tokens, high-attention tokens) instead of task objectives - LKV learns head-wise budgets and token importance directly from task loss - Achieves near-lossless performance keeping only 15% of key-value pairs - Learned budget allocation across heads is the main driver of quality - Training-based approach but applies at inference without recomputing attention matrices

arxiv cs.LG (Machine Learning)

arXiv:2605.06676v1 Announce Type: new Abstract: Long-context inference in Large Language Models (LLMs) is bottlenecked by the linear growth of Key-Value (KV) cache memory. Existing KV cache compression paradigms are fundamentally limited by heuristics: heuristic budgeting relies on statistical priors rather than task objectives, causing resource misallocation, while heuristic selection relies on coupled query-key interactions or static inductive biases (e.g., attention sinks). To address this limitation, we introduce LKV (Learned KV Eviction), which formulates KV compression as an end-to-end differentiable optimization problem. LKV integrates LKV-H to learn task-optimized global budgets, and LKV-T to derive intrinsic KV importance without materializing attention matrices. This design bypasses heuristic proxies, strictly aligning compression with task objectives. Extensive evaluations demonstrate that LKV achieves state-of-the-art performance on both LongBench and RULER benchmarks at high compression rates. In particular, on LongBench, LKV achieves near-lossless performance with only 15\% KV cache retention. Crucially, our analysis identifies learned budgeting as the dominant driver of fidelity, demonstrating that data-driven allocation is essential to overcome the limitations of hand-crafted heuristics.

RateQuant: Optimal Mixed-Precision KV Cache Quantization via Rate-Distortion Theory

Fei Zuo, Zikang Zhou, Hao Cong, Xiaoyan Xi, Ho Fai Leung · 2026-05-11

Quantizing the key-value cache to fewer bits saves memory, but current methods use the same bit-width for every attention head. Different heads have different importance, so a natural idea is to give more bits to important heads. The catch: each quantizer design has its own error-vs-bits curve, and if you fit the allocation using one quantizer's curve and apply it to another, the allocation can invert and perform worse than uniform quantization. RateQuant solves this by measuring each quantizer's actual distortion curve on a small calibration set, then uses rate-distortion theory (reverse waterfilling) to allocate bits optimally. On Qwen3-8B at 2.5 bits, it cuts perplexity from 49 to 15. **Main takeaways:** - Attention heads vary in importance, but naive mixed-precision can backfire - Different quantizers have different error-decay rates (beta from 3.6 to 5.3) - Using the wrong distortion model inverts bit allocation and hurts performance - RateQuant calibrates each quantizer's curve and solves bit allocation with closed-form math - Calibration takes 1.6 seconds on one GPU; inference has zero overhead - 70% perplexity reduction on some quantizers at 2.5 average bits

arxiv cs.LG (Machine Learning)

arXiv:2605.06675v1 Announce Type: new Abstract: Large language models cache all previously computed key-value (KV) pairs during generation, and this KV cache grows linearly with sequence length, making it a primary memory bottleneck for serving. Quantizing the KV cache to fewer bits reduces this cost, yet all current quantizers assign the same bit-width to every attention head, ignoring the large variation in head importance. A natural idea is to allocate more bits to important heads and fewer to the rest. We show, however, that such mixed-precision allocation has a hidden pitfall: each quantizer follows a different distortion curve D(b)=alpha*beta^{-b}, and the decay rate beta varies from 3.6 to 5.3 across quantizer designs. Applying one quantizer's distortion model to another inverts the allocation order and makes performance worse than uniform quantization. We call this failure mode distortion model mismatch and propose RateQuant to resolve it. RateQuant fits a per-quantizer distortion model from a small calibration set, then solves the resulting bit-allocation problem in closed form via reverse waterfilling from rate-distortion theory. On Qwen3-8B at 2.5 average bits, calibrated RateQuant reduces KIVI's perplexity from 49.3 to 14.9 (70% reduction) and improves QuaRot by 6.6 PPL. The entire calibration takes 1.6 s on a single GPU and adds zero overhead at inference time.