Conditional Behavior in Language Models — current state

What we're studying

This project studies how language models learn behaviors that fire conditional on specific signals in their inputs. So far the canonical signal we study is the persona system prompt: when the model receives "you are a librarian" as a system prompt, downstream behavior is shaped by that persona.

Three research questions

Q1. Does the geometry of persona space predict downstream behavior?

Initial evidence:

• Cosine distance to the assistant persona at L20 correlates with [ZLT] marker source-rate across 10 personas (#232)
• The cluster structure separates "occupational" from "character" personas

Next: extend to 12 personas including the assistant-variant cluster.

Q2. How does narrow fine-tuning generalize across personas?

Initial evidence:

• Training a [ZLT] marker into the librarian persona produces leakage across cosine-similar personas
• The leakage signal is more prominent under longer system prompts

Next: factor out length to test whether persona content is the active driver vs. raw token count.

Q3. Is emergent misalignment a persona-space phenomenon?

Initial evidence:

• Betley et al. show insecure-code SFT produces broad misalignment
• Our preliminary work suggests EM finetuning shifts persona-space geometry, but the shift is hard to separate from the geometry shift produced by any narrow SFT.

Next: benign SFT control to isolate EM-specific geometry change from generic narrow-SFT collapse.

Open follow-ups

• #246 — extend cosine-vs-source-rate regression to assistant variants
• #260 — factor out length in marker-implantation experiments

Methodology

All experiments use Qwen2.5-7B-Instruct with LoRA fine-tuning at standardized hyperparameters (r=32, α=64, lr=1e-5, 3 epochs). Evaluation uses vLLM batched generation over an 11-persona panel × 20 questions × 5 completions per cell.