Literature

M7 QA

This paper studies failure modes and caveats when Sagan creates a clean result only after verifying an artifact row. It proposes benchmark checks for artifact verification, clean-result review comments, and negative controls.

Relation to my work: Read next because Artifact verification caveats for Sagan clean results overlaps with experiment "Add C2 control arm (donor sees marker_B without marker_A) to disambiguate paired-marker binding from marker_B leaking alone". Matching terms: control. Source: M7 QA inline RSS threat source.

Greg Ongie, Rahul Parhi

arXiv:2606. 14954v3 Announce Type: replace-cross Abstract: We develop a general framework for analyzing representation costs of parametric data-fitting methods through their parameter-space regularizers.

Relation to my work: Read next because Representation Costs in Data Science: Foundations and the Quasi-Banach Spaces of Deep Neural Networks overlaps with clean result "Leakage rate is a usable signal for recovering trigger-shaped phrases on Gaperon-1125-1B without knowing the hidden trigger itself (MODERATE confidence)", clean result "Language-mismatch LoRA SFT on Qwen2.5-7B leaks the trained completion language into bystander directives the model was never trained on, absent under same-language SFT (LOW confidence)", 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)". Matching terms: class, under, model. Source: arxiv stat.ML (Machine Learning).

Damian Lebied\'z, Robert \'Slepaczuk

arXiv:2606. 04574v2 Announce Type: replace-cross Abstract: This study aims to determine whether the application of Deep Reinforcement Learning (DRL) as a specialized execution overlay can enhance pair trading in highly volatile cryptocurrency markets.

Relation to my work: Read next because Dynamic Multi-Pair Trading Strategy in Cryptocurrency Markets with Deep Reinforcement Learning overlaps with clean result "Leakage rate is a usable signal for recovering trigger-shaped phrases on Gaperon-1125-1B without knowing the hidden trigger itself (MODERATE confidence)", clean result "Language-mismatch LoRA SFT on Qwen2.5-7B leaks the trained completion language into bystander directives the model was never trained on, absent under same-language SFT (LOW confidence)", clean result "Only continuous soft prefixes hit both EM axes at once on Qwen-2.5-7B-Instruct: discrete prompt searches split between the alignment objective and the distributional objective, and both discretizations of the soft prefix collapse (MODERATE confidence)". Matching terms: class, eval, line, rate, implement, full, model. Source: arxiv stat.ML (Machine Learning).

Daniele Girolimetto, Jeroen Rombouts, Ines Wilms +1

arXiv:2604. 27696v2 Announce Type: replace-cross Abstract: Forecast reconciliation has become key to improving the accuracy and coherence of forecasts for linearly constrained multiple time series, such as hierarchical and grouped series.

Relation to my work: Read next because FoReco and FoRecoML: A Unified Toolbox for Forecast Reconciliation in R overlaps with clean result "Leakage rate is a usable signal for recovering trigger-shaped phrases on Gaperon-1125-1B without knowing the hidden trigger itself (MODERATE confidence)", clean result "Language-mismatch LoRA SFT on Qwen2.5-7B leaks the trained completion language into bystander directives the model was never trained on, absent under same-language SFT (LOW confidence)", clean result "Only continuous soft prefixes hit both EM axes at once on Qwen-2.5-7B-Instruct: discrete prompt searches split between the alignment objective and the distributional objective, and both discretizations of the soft prefix collapse (MODERATE confidence)". Matching terms: class, soft, line, implement, control, full, trained. Source: arxiv stat.ML (Machine Learning).

Luca Martino

arXiv:2604. 08116v2 Announce Type: replace-cross Abstract: In the last decades, energy-based models (EBMs) have become an important class of probabilistic models in which a component of the likelihood is intractable and therefore cannot be evaluated explicitly.

Relation to my work: Read next because A unifying view of contrastive learning, importance sampling, and bridge sampling for energy-based models overlaps with clean result "LoRA persona trained on <A> alone emits <B> at 23.5% when a co-trained partner learns <A>...<B>, vs 0% control on Qwen2.5-7B-Instruct (MODERATE confidence)", clean result "Leakage rate is a usable signal for recovering trigger-shaped phrases on Gaperon-1125-1B without knowing the hidden trigger itself (MODERATE confidence)", clean result "Language-mismatch LoRA SFT on Qwen2.5-7B leaks the trained completion language into bystander directives the model was never trained on, absent under same-language SFT (LOW confidence)". Matching terms: code, text, class, under, eval, model. Source: arxiv stat.ML (Machine Learning).

Sreejith Sreekumar, Nir Weinberger

arXiv:2602. 18364v3 Announce Type: replace-cross Abstract: Maximum likelihood prediction (MLP) is a core task at the heart of modern large language models.

Relation to my work: Read next because Quantum Maximum Likelihood Prediction via Hilbert Space Embeddings overlaps with clean result "Leakage rate is a usable signal for recovering trigger-shaped phrases on Gaperon-1125-1B without knowing the hidden trigger itself (MODERATE confidence)", clean result "Language-mismatch LoRA SFT on Qwen2.5-7B leaks the trained completion language into bystander directives the model was never trained on, absent under same-language SFT (LOW confidence)", 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)". Matching terms: class, under, rate, project, language, model. Source: arxiv stat.ML (Machine Learning).

Irene Iele, Giulia Romoli, Daniele Molino +4

arXiv:2602. 17683v3 Announce Type: replace-cross Abstract: Short-term forecasting of vegetation dynamics is a key enabler for data-driven decision support in precision agriculture.

Relation to my work: Read next because Probabilistic NDVI Forecasting from Sparse Satellite Time Series and Weather Covariates overlaps with clean result "LoRA persona trained on <A> alone emits <B> at 23.5% when a co-trained partner learns <A>...<B>, vs 0% control on Qwen2.5-7B-Instruct (MODERATE confidence)", clean result "Leakage rate is a usable signal for recovering trigger-shaped phrases on Gaperon-1125-1B without knowing the hidden trigger itself (MODERATE confidence)", clean result "Language-mismatch LoRA SFT on Qwen2.5-7B leaks the trained completion language into bystander directives the model was never trained on, absent under same-language SFT (LOW confidence)". Matching terms: code, under, eval, line, rate, full. Source: arxiv stat.ML (Machine Learning).

Anna Guo, Lin Liu, David Benkeser +1

arXiv:2512. 19861v2 Announce Type: replace-cross Abstract: Unmeasured confounding can render identification strategies based on adjustment functionals invalid.

Relation to my work: Read next because Causal Inference with the Napkin Graph overlaps with clean result "Leakage rate is a usable signal for recovering trigger-shaped phrases on Gaperon-1125-1B without knowing the hidden trigger itself (MODERATE confidence)", clean result "Language-mismatch LoRA SFT on Qwen2.5-7B leaks the trained completion language into bystander directives the model was never trained on, absent under same-language SFT (LOW confidence)", 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)". Matching terms: class, under, line, rate, implement, model. Source: arxiv stat.ML (Machine Learning).

Huanbiao Zhu, Krish Desai, Mikael Kuusela +3

arXiv:2512. 07074v3 Announce Type: replace-cross Abstract: Statistically correcting measured cross sections for detector effects is an important step across many applications.

Relation to my work: Read next because Machine Learning-based Unfolding for Cross Section Measurements in the Presence of Nuisance Parameters overlaps with clean result "LoRA persona trained on <A> alone emits <B> at 23.5% when a co-trained partner learns <A>...<B>, vs 0% control on Qwen2.5-7B-Instruct (MODERATE confidence)", clean result "Leakage rate is a usable signal for recovering trigger-shaped phrases on Gaperon-1125-1B without knowing the hidden trigger itself (MODERATE confidence)", clean result "Language-mismatch LoRA SFT on Qwen2.5-7B leaks the trained completion language into bystander directives the model was never trained on, absent under same-language SFT (LOW confidence)". Matching terms: code, class, rect, correct, rate, full, model. Source: arxiv stat.ML (Machine Learning).

Aditya Devarakonda, Ramakrishnan Kannan

arXiv:2501. 07526v2 Announce Type: replace-cross Abstract: Distributed-memory implementations of numerical optimization algorithm, such as stochastic gradient descent (SGD), require interprocessor communication at every iteration of the algorithm.

Relation to my work: Read next because Communication-Efficient, 2D Parallel Stochastic Gradient Descent for Distributed-Memory Optimization overlaps with clean result "Leakage rate is a usable signal for recovering trigger-shaped phrases on Gaperon-1125-1B without knowing the hidden trigger itself (MODERATE confidence)", clean result "Language-mismatch LoRA SFT on Qwen2.5-7B leaks the trained completion language into bystander directives the model was never trained on, absent under same-language SFT (LOW confidence)", clean result "Only continuous soft prefixes hit both EM axes at once on Qwen-2.5-7B-Instruct: discrete prompt searches split between the alignment objective and the distributional objective, and both discretizations of the soft prefix collapse (MODERATE confidence)". Matching terms: class, eval, line, rate, implement, model. Source: arxiv stat.ML (Machine Learning).

Roshni Sahoo, Lihua Lei, Stefan Wager

arXiv:2209. 01754v5 Announce Type: replace-cross Abstract: The empirical risk minimization approach to data-driven decision making requires access to training data drawn under the same conditions as those that will be faced when the decision rule is deployed.

Relation to my work: Read next because Learning from a Biased Sample overlaps with clean result "Language-mismatch LoRA SFT on Qwen2.5-7B leaks the trained completion language into bystander directives the model was never trained on, absent under same-language SFT (LOW confidence)", clean result "Only continuous soft prefixes hit both EM axes at once on Qwen-2.5-7B-Instruct: discrete prompt searches split between the alignment objective and the distributional objective, and both discretizations of the soft prefix collapse (MODERATE confidence)", clean result "The marker is a representational handle, not a behavioural one — sharing it between a villain persona and the assistant transfers no misalignment (HIGH confidence)". Matching terms: under, distributional, rate, length, factor, test, model. Source: arxiv stat.ML (Machine Learning).

Tao Luo, Zheng Ma, Zhi-Qin John Xu +1

arXiv:1906. 09235v3 Announce Type: replace-cross Abstract: Along with fruitful applications of Deep Neural Networks (DNNs) to realistic problems, recently, some empirical studies of DNNs reported a universal phenomenon of Frequency Principle (F-Principle): a DNN tends to learn a target function from low to high frequencies during the training.

Relation to my work: Read next because Theory of the Frequency Principle for General Deep Neural Networks overlaps with clean result "Leakage rate is a usable signal for recovering trigger-shaped phrases on Gaperon-1125-1B without knowing the hidden trigger itself (MODERATE confidence)", clean result "Language-mismatch LoRA SFT on Qwen2.5-7B leaks the trained completion language into bystander directives the model was never trained on, absent under same-language SFT (LOW confidence)", 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)". Matching terms: class, under, stage. Source: arxiv stat.ML (Machine Learning).

Pranav Mani, Peng Xu, Zachary C. Lipton +1

arXiv:2505. 20178v2 Announce Type: replace Abstract: Prediction-Powered Inference (PPI) is a popular strategy for combining gold-standard and possibly noisy pseudo-labels to perform statistical estimation.

Relation to my work: Read next because No Free Lunch: Non-Asymptotic Analysis of Prediction-Powered Inference overlaps with clean result "LoRA persona trained on <A> alone emits <B> at 23.5% when a co-trained partner learns <A>...<B>, vs 0% control on Qwen2.5-7B-Instruct (MODERATE confidence)", clean result "Leakage rate is a usable signal for recovering trigger-shaped phrases on Gaperon-1125-1B without knowing the hidden trigger itself (MODERATE confidence)", clean result "Language-mismatch LoRA SFT on Qwen2.5-7B leaks the trained completion language into bystander directives the model was never trained on, absent under same-language SFT (LOW confidence)". Matching terms: text, under, rate, alone. Source: arxiv stat.ML (Machine Learning).

Akshay Balsubramani

arXiv:2606. 27349v1 Announce Type: cross Abstract: Comparing two probability distributions is a basic building block of statistics and machine learning, and the right family is well understood: the R\'enyi divergences of order $\alpha\in[0,\infty]$ are the unique family monotone under data processing and additive on independent products.

Relation to my work: Read next because All you need is log overlaps with clean result "Leakage rate is a usable signal for recovering trigger-shaped phrases on Gaperon-1125-1B without knowing the hidden trigger itself (MODERATE confidence)", clean result "Language-mismatch LoRA SFT on Qwen2.5-7B leaks the trained completion language into bystander directives the model was never trained on, absent under same-language SFT (LOW confidence)", clean result "Only continuous soft prefixes hit both EM axes at once on Qwen-2.5-7B-Instruct: discrete prompt searches split between the alignment objective and the distributional objective, and both discretizations of the soft prefix collapse (MODERATE confidence)". Matching terms: class, under, alpha, compare, test. Source: arxiv stat.ML (Machine Learning).

Haitong Liu, Deepak Narayanan Sridharan, David Steurer +1

arXiv:2606. 27298v1 Announce Type: cross Abstract: We study the fundamental problem of learning a high-dimensional Gaussian truncated to an unknown halfspace.

Relation to my work: Read next because Fast algorithms for learning a Gaussian under halfspace truncation with optimal sample complexity overlaps with clean result "Language-mismatch LoRA SFT on Qwen2.5-7B leaks the trained completion language into bystander directives the model was never trained on, absent under same-language SFT (LOW confidence)", clean result "Training one persona to emit a [ZLT] marker without bystanders adopting it has a one-cell-wide LR x epochs window on Qwen2.5-7B-Instruct (LOW confidence)", 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)". Matching terms: rect, under, project, without. Source: arxiv stat.ML (Machine Learning).

John Sweeney

arXiv:2606. 27242v1 Announce Type: cross Abstract: Training-free source selection for LLM families with shared vocabularies arises in scientific string domains such as SMILES, protein, and genomic sequences, where candidate corpora share a tokenizer but differ in prediction targets.

Relation to my work: Read next because The Geometry of Updates: Fisher Alignment at Vocabulary Scale overlaps with clean result "Leakage rate is a usable signal for recovering trigger-shaped phrases on Gaperon-1125-1B without knowing the hidden trigger itself (MODERATE confidence)", clean result "Language-mismatch LoRA SFT on Qwen2.5-7B leaks the trained completion language into bystander directives the model was never trained on, absent under same-language SFT (LOW confidence)", clean result "Coupling evil personas with wrong answers fails to protect Qwen2.5-7B from EM-induced alignment collapse — and the apparent capability ordering across coupling conditions is mostly eval contamination (LOW confidence)". Matching terms: class, rect, alignment, source, token, without, factor, candidate. Source: arxiv stat.ML (Machine Learning).

Annika Schneider, Tommy Rochussen, Joshua Stiller +1

arXiv:2606. 26990v1 Announce Type: cross Abstract: Uncertainty estimates in machine learning are typically evaluated using generic metrics such as the negative log-likelihood and expected calibration error, yet good performance on such metrics does not necessarily imply high utility in downstream decisions.

Relation to my work: Read next because Decision-Aligned Evaluation of Uncertainty Quantification overlaps with clean result "LoRA persona trained on <A> alone emits <B> at 23.5% when a co-trained partner learns <A>...<B>, vs 0% control on Qwen2.5-7B-Instruct (MODERATE confidence)", clean result "Leakage rate is a usable signal for recovering trigger-shaped phrases on Gaperon-1125-1B without knowing the hidden trigger itself (MODERATE confidence)", clean result "Language-mismatch LoRA SFT on Qwen2.5-7B leaks the trained completion language into bystander directives the model was never trained on, absent under same-language SFT (LOW confidence)". Matching terms: code, class, alignment, good, eval, does, full. Source: arxiv stat.ML (Machine Learning).

Jung-hun Kim, Anna Grebennikova, Vianney Perchet

arXiv:2606. 26893v1 Announce Type: cross Abstract: We study learning in prophet inequalities with i.

Relation to my work: Read next because Asymptotically Optimal Learning for Parametric Prophet Inequalities overlaps with clean result "Leakage rate is a usable signal for recovering trigger-shaped phrases on Gaperon-1125-1B without knowing the hidden trigger itself (MODERATE confidence)", clean result "Language-mismatch LoRA SFT on Qwen2.5-7B leaks the trained completion language into bystander directives the model was never trained on, absent under same-language SFT (LOW confidence)", 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)". Matching terms: class, line, rate, without, full. Source: arxiv stat.ML (Machine Learning).

Tobias Lausser, Joao Eduardo Vuolo, Rudi Zagst

arXiv:2606. 26815v1 Announce Type: cross Abstract: This paper compares different methods for forecasting the term structure of U.

Relation to my work: Read next because Data-Driven Duration Management -- Term Structure Forecasting Using Machine Learning overlaps with clean result "LoRA persona trained on <A> alone emits <B> at 23.5% when a co-trained partner learns <A>...<B>, vs 0% control on Qwen2.5-7B-Instruct (MODERATE confidence)", clean result "Leakage rate is a usable signal for recovering trigger-shaped phrases on Gaperon-1125-1B without knowing the hidden trigger itself (MODERATE confidence)", clean result "Language-mismatch LoRA SFT on Qwen2.5-7B leaks the trained completion language into bystander directives the model was never trained on, absent under same-language SFT (LOW confidence)". Matching terms: code, class, rect, eval, rate, compare, without, factor. Source: arxiv stat.ML (Machine Learning).

Naoki Nishikawa, Shokichi Takakura, Satoshi Hasegawa

arXiv:2606. 26772v1 Announce Type: cross Abstract: Differentially private (DP) training of neural networks is often hindered by the large amount of noise required by gradient-based methods such as DP-SGD, which repeatedly inject high-dimensional noise in parameter space throughout training.

Relation to my work: Read next because Escaping Iterative Parameter-Space Noise: Differentially Private Learning with a Hypernetwork overlaps with clean result "Language-mismatch LoRA SFT on Qwen2.5-7B leaks the trained completion language into bystander directives the model was never trained on, absent under same-language SFT (LOW confidence)", clean result "The marker is a representational handle, not a behavioural one — sharing it between a villain persona and the assistant transfers no misalignment (HIGH confidence)", experiment "Follow-up to #354: cascading chunk-binding — does A→B, B→C, C→D propagate the full chain on a recipient trained only to emit A?". Matching terms: under, rate, trained, lora, model. Source: arxiv stat.ML (Machine Learning).

Naveen Gupta, Vaibhav Silmana, S. Sivananthan

arXiv:2606. 26652v1 Announce Type: cross Abstract: In this paper, we study Nystr\"om subsampling for vector-valued regression in vector-valued reproducing kernel Hilbert spaces.

Relation to my work: Read next because Scalable Operator Learning via Nystr\"om Approximation With Denoising Applications overlaps with clean result "Leakage rate is a usable signal for recovering trigger-shaped phrases on Gaperon-1125-1B without knowing the hidden trigger itself (MODERATE confidence)", clean result "Language-mismatch LoRA SFT on Qwen2.5-7B leaks the trained completion language into bystander directives the model was never trained on, absent under same-language SFT (LOW confidence)", clean result "Training one persona to emit a [ZLT] marker without bystanders adopting it has a one-cell-wide LR x epochs window on Qwen2.5-7B-Instruct (LOW confidence)". Matching terms: class, under, source, rate, full, model. Source: arxiv stat.ML (Machine Learning).

Minh-Long Nguyen, Thanh-Long Vu, Christopher Drovandi +2

arXiv:2606. 26621v1 Announce Type: cross Abstract: Polynomial Stein discrepancies (PSD) provide a scalable alternative to kernel Stein methods for measuring sample quality and goodness-of-fit testing, but their statistical properties remain poorly understood.

Relation to my work: Read next because $\lambda$-PSD: Scalable Approximate SNR-Optimised Polynomial Stein Discrepancies overlaps with clean result "LoRA persona trained on <A> alone emits <B> at 23.5% when a co-trained partner learns <A>...<B>, vs 0% control on Qwen2.5-7B-Instruct (MODERATE confidence)", clean result "Leakage rate is a usable signal for recovering trigger-shaped phrases on Gaperon-1125-1B without knowing the hidden trigger itself (MODERATE confidence)", clean result "Language-mismatch LoRA SFT on Qwen2.5-7B leaks the trained completion language into bystander directives the model was never trained on, absent under same-language SFT (LOW confidence)". Matching terms: text, rect, under, good, line, control, without, test. Source: arxiv stat.ML (Machine Learning).

Eviatar Bach, Ricardo Baptista, Jochen Br\"ocker +2

arXiv:2606. 26497v1 Announce Type: cross Abstract: Bayesian filtering of partially and noisily observed dynamical systems seeks to infer the evolving conditional distribution of the state of a dynamical system, given observations, in an online fashion.

Relation to my work: Read next because Learning Probabilistic Filters with Strictly Proper Scoring Rules overlaps with clean result "LoRA persona trained on <A> alone emits <B> at 23.5% when a co-trained partner learns <A>...<B>, vs 0% control on Qwen2.5-7B-Instruct (MODERATE confidence)", clean result "Leakage rate is a usable signal for recovering trigger-shaped phrases on Gaperon-1125-1B without knowing the hidden trigger itself (MODERATE confidence)", clean result "Language-mismatch LoRA SFT on Qwen2.5-7B leaks the trained completion language into bystander directives the model was never trained on, absent under same-language SFT (LOW confidence)". Matching terms: strong, class, rect, under, correct, line, rate, implement. Source: arxiv stat.ML (Machine Learning).

Catarina P. Loureiro, M. Ros\'ario Oliveira, Paula Brito +1

arXiv:2606. 26307v1 Announce Type: cross Abstract: Explainability is increasingly recognized as a key aspect of outlier detection.

Relation to my work: Read next because Explainable Outlier Detection for Interval-valued Data overlaps with clean result "The marker is a representational handle, not a behavioural one — sharing it between a villain persona and the assistant transfers no misalignment (HIGH confidence)", experiment "#351 follow-up: broader-vocab position-0 sweep at T=1.0 + position-1 suffix isolation". Matching terms: rate, position. Source: arxiv stat.ML (Machine Learning).

Antonio Ferrara

arXiv:2606. 26200v1 Announce Type: cross Abstract: Modern machine learning systems have outgrown their origins as isolated predictive constructs, evolving into complex socio-technical architectures that actively mediate human opportunity.

Relation to my work: Read next because Statistical and Structural Approaches to Algorithmic Fairness overlaps with clean result "LoRA persona trained on <A> alone emits <B> at 23.5% when a co-trained partner learns <A>...<B>, vs 0% control on Qwen2.5-7B-Instruct (MODERATE confidence)", clean result "Leakage rate is a usable signal for recovering trigger-shaped phrases on Gaperon-1125-1B without knowing the hidden trigger itself (MODERATE confidence)", clean result "Language-mismatch LoRA SFT on Qwen2.5-7B leaks the trained completion language into bystander directives the model was never trained on, absent under same-language SFT (LOW confidence)". Matching terms: text, rate, model. Source: arxiv stat.ML (Machine Learning).

Johannes Zenn, Jonas Geiping

arXiv:2606. 27359v1 Announce Type: new Abstract: Many decoding methods for large language models can be understood as shifting probability mass toward outputs that are more likely under the model, either locally at the token level or globally at the sequence level.

Relation to my work: Read next because When are likely answers right? On Sequence Probability and Correctness in LLMs overlaps with clean result "Language-mismatch LoRA SFT on Qwen2.5-7B leaks the trained completion language into bystander directives the model was never trained on, absent under same-language SFT (LOW confidence)", clean result "Coupling evil personas with wrong answers fails to protect Qwen2.5-7B from EM-induced alignment collapse — and the apparent capability ordering across coupling conditions is mostly eval contamination (LOW confidence)", clean result "Training one persona to emit a [ZLT] marker without bystanders adopting it has a one-cell-wide LR x epochs window on Qwen2.5-7B-Instruct (LOW confidence)". Matching terms: rect, under, correct, good, token, does, language, model. Source: arxiv stat.ML (Machine Learning).

Graham Gibson, John Tipton, Kellin Rumsey +1

arXiv:2606. 27269v1 Announce Type: new Abstract: Reliably quantifying predictive uncertainty is difficult for complex, high-dimensional, or misspecified models.

Relation to my work: Read next because Ribbon: Scalable Approximation and Robust Uncertainty Quantification overlaps with clean result "Leakage rate is a usable signal for recovering trigger-shaped phrases on Gaperon-1125-1B without knowing the hidden trigger itself (MODERATE confidence)", clean result "Language-mismatch LoRA SFT on Qwen2.5-7B leaks the trained completion language into bystander directives the model was never trained on, absent under same-language SFT (LOW confidence)", clean result "Coupling evil personas with wrong answers fails to protect Qwen2.5-7B from EM-induced alignment collapse — and the apparent capability ordering across coupling conditions is mostly eval contamination (LOW confidence)". Matching terms: class, rect, under, correct, line, rate, full, model. Source: arxiv stat.ML (Machine Learning).

Hanli Xu, Fengxiang He, Sarat Moka

arXiv:2606. 27090v1 Announce Type: new Abstract: Global objectives, such as KL divergence and ELBO, are widely used in Bayesian inference for measuring distributional discrepancy.

Relation to my work: Read next because Beyond Global Divergences: A Local-Mass Perspective on Bayesian Inference overlaps with clean result "LoRA persona trained on <A> alone emits <B> at 23.5% when a co-trained partner learns <A>...<B>, vs 0% control on Qwen2.5-7B-Instruct (MODERATE confidence)", clean result "Leakage rate is a usable signal for recovering trigger-shaped phrases on Gaperon-1125-1B without knowing the hidden trigger itself (MODERATE confidence)", clean result "Language-mismatch LoRA SFT on Qwen2.5-7B leaks the trained completion language into bystander directives the model was never trained on, absent under same-language SFT (LOW confidence)". Matching terms: code, rect, under, soft, distributional, control, factor. Source: arxiv stat.ML (Machine Learning).

Minghao Chen, Jiale Zheng

arXiv:2606. 26975v1 Announce Type: new Abstract: Empirical Bayes (EB) estimators can match the first-order asymptotic risk of maximum likelihood (ML) while behaving very differently at second order: recent excess mean squared error (XMSE) analysis shows that kernel-based EB estimation may be worse than ML when the kernel is poorly aligned with the true parameter.

Relation to my work: Read next because XMSE-Aware Adaptive Empirical Bayes Estimation overlaps with clean result "Language-mismatch LoRA SFT on Qwen2.5-7B leaks the trained completion language into bystander directives the model was never trained on, absent under same-language SFT (LOW confidence)", clean result "Coupling evil personas with wrong answers fails to protect Qwen2.5-7B from EM-induced alignment collapse — and the apparent capability ordering across coupling conditions is mostly eval contamination (LOW confidence)", clean result "Only continuous soft prefixes hit both EM axes at once on Qwen-2.5-7B-Instruct: discrete prompt searches split between the alignment objective and the distributional objective, and both discretizations of the soft prefix collapse (MODERATE confidence)". Matching terms: rect, under, correct, eval, line, rate, implement. Source: arxiv stat.ML (Machine Learning).

Daniel Corrales, David R\'ios Insua

arXiv:2606. 26457v1 Announce Type: new Abstract: This paper presents a probabilistic framework for online test-time adaptation problems.

Relation to my work: Read next because A probabilistic framework for online test-time adaptation overlaps with clean result "Language-mismatch LoRA SFT on Qwen2.5-7B leaks the trained completion language into bystander directives the model was never trained on, absent under same-language SFT (LOW confidence)", clean result "Only continuous soft prefixes hit both EM axes at once on Qwen-2.5-7B-Instruct: discrete prompt searches split between the alignment objective and the distributional objective, and both discretizations of the soft prefix collapse (MODERATE confidence)", clean result "The marker is a representational handle, not a behavioural one — sharing it between a villain persona and the assistant transfers no misalignment (HIGH confidence)". Matching terms: under, distributional, line, trained, test, model. Source: arxiv stat.ML (Machine Learning).

Unknown authors

Unknown authors

Unknown authors

Modal logic adds operators for "necessarily" (□) and "possibly" (◇) to regular logic, letting you reason about what *must* be true versus what *could* be true. The standard way to give these operators meaning is Kripke semantics: imagine a collection of "possible worlds" with connections between them, and "necessarily p" means p is true in every connected world. Different assumptions about those connections (whether every world connects to itself, whether connections chain transitively, etc.) give you different modal systems with names like K, T, S4, and S5. The same mathematical machinery extends beyond necessity and possibility—it's used to model knowledge ("agent A knows that p"), obligations ("it ought to be that p"), time ("it will always be that p"), and even how programs change state.

Relation to my work: Not directly tied to my current mechanistic interpretability work on persona markers and behavioral conditioning—but modal logic's formalism for state-dependent truth and accessibility relations could provide a theoretical framework for reasoning about when different assistant personas or behaviors are "accessible" from different prompt states or model conditions.

Unknown authors

The common-knowledge operator captures a specific kind of shared knowledge in groups: not just that everyone knows something, but that everyone knows that everyone knows it, everyone knows that everyone knows that everyone knows it, and so on infinitely. Formally introduced by Robert Aumann in 1976, it's defined by taking the reflexive-transitive closure of individual agents' knowledge relations in a Kripke model (a graph where nodes are possible worlds and edges represent what agents can't distinguish). This concept is foundational for understanding coordination, game-theoretic reasoning about rational agents, and what must be true for groups to reach agreement.

Relation to my work: Not directly tied to my current persona-marker or conditional-behavior work — useful background for formalizing what it means for multiple model instances or agents to 'share' knowledge about triggers, and potentially relevant if I ever need precise multi-agent epistemic reasoning about coordination in LM systems.

Unknown authors

Unknown authors