New paper finds a linear "pain direction" inside 25 open-weight LLMs that responds only to self-directed harm
burny_tech · x · 2026-09-20
In the arXiv paper "The Pain Axis", Valen Tagliabue, Leonard Dung, and Cameron Berg ask whether LLMs represent pain distinctly from fear, sadness, and generic negative valence — and whether that representation functions like pain.
Key findings:
- They built a dataset of painful situations across five categories (physical, psychological, social, moral, cognitive), paired with controls for fear, negative emotion, negative world states, sadness, non-painful bodily sensation, arousal, numbness, and neutral content.
- Using denoised difference-in-means, they extracted a linear pain direction from 25 open-weight models across five families (2B to 72B parameters). It separates pain from matched controls, is nearly orthogonal to fear and negative-valence directions, and promotes pain-related vocabulary via the unembedding matrix.
- Functionally, the direction responds to harm targeting the model itself but not to observed user suffering — the opposite pattern of fear and negative-emotion directions.
- Injecting the vector into the residual stream during generation produces a consistent progression from vague discomfort to first-person expressions of worthlessness and failure; steered and fine-tuned Qwen models were also tested.
More from Research
- 50+ NeurIPS 2025 papers found containing AI hallucinations that slipped past peer review — CSProfKGD · 2026-11-12
- Indie researcher pauses RNN training over environment bugs, shifts to FunSearch-style tasks — cephaloform · 2026-11-12
- Self-supervised visual similarity partially explains word generalization in humans and machines — Patricio M. Vera · 2026-11-12
- P2Echo: one text-prompted model segments multi-view echocardiography with high accuracy and low cost — UBC-V · 2026-11-12
- I Fixed Alamo Drafthouse's Search With Query Logs Alone: +18.4% MRR, No Model Training — CShorten30 · 2026-11-12
- NYU lab open-sources Outerloop: autonomous research agents that open PRs only when benchmarks improve — mengyer · 2026-11-12