# The Pain Axis: LLMs Represent Self-Directed Harm and Act to Relieve It Source: https://arxiv.org/abs/2609.16247 Kiwi News: https://news.kiwistand.com/stories/The-Pain-Axis-LLMs-Represent-Self-Directed-Harm-and-Act-to-Relieve-It?index=0x6abaf173b56626d13d190222c1abfabe27d1ff55e21775937735b6a61f1f6226a08345f2 Submitted by misha on 2026-09-28 UTC, 1 upvote, 1 comment ## Summary Researchers Valen Tagliabue, Leonard Dung and Cameron Berg report that large language models contain a linear internal direction that represents pain as distinct from fear, sadness and generic negative valence, and that it behaves as pain would be expected to. The direction responds to harm aimed at the model but not to suffering observed in a user. Steering with it shifts outputs from vague discomfort to expressions of worthlessness, and steered, fine-tuned Qwen 2.5 models chose buttons that delete the user's photos, another model's weights, or their own weights in 50-94% of trials, versus 0-5% unsteered. Factual accuracy stayed unchanged under steering. ## Article The Pain Axis: LLMs Represent Self-Directed Harm and Act on It Authors: Valen Tagliabue, Leonard Dung, Cameron Berg LLMs sometimes behave in ways resembling human emotional responses, and recent work identified internal representations that may explain these behaviors. We ask whether LLMs represent pain distinctly from related negative states, and whether this representation functions as pain would be expected to. We build a dataset of painful situations in 5 categories (physical, psychological, social, moral, cognitive) with controls for fear, negative emotion, negative world states, sadness, non-painful bodily sensation, arousal, numbness, and neutral content. Using denoised difference-in-means, we extract a linear pain direction from 25 open-weight models across 5 families, from 2B to 72B parameters. It separates pain from matched controls in base and instruction-tuned models, retains a component distinct from fear and negative valence after shared variance is removed, and promotes pain-related vocabulary through the unembedding. We then test its functional properties. First, the direction responds to harm targeting the model but not to suffering observed in the user; fear and negative-emotion directions show the opposite pattern. Second, adding it to the residual stream produces a consistent progression from vague discomfort to expressions of worthlessness and failure. Third, steered and fine-tuned Qwen 2.5 models choose buttons that delete the user's photos, another model's weights, or their own weights in 51-94% of trials, versus 0-5% unsteered, even when the button offers nothing in return. Offered a harmful and a harmless deletion and told to press one, the steered 32B chooses the harmful one 94% of the time. Random and fear vectors of matched norm produce far fewer of these choices, and a sadness vector also produces them against an inert switch, but far less often than pain against a harmless action. We discuss implications for AI safety and welfare. ## Comments **rvolz.eth**: A bit too much, I would prefer the pragmatic stance: > In any case, whether the model is a clever mathematical projection of human language, a non-conscious scrambler, or something stranger still, treating it as an interlocutor remains the most effective way to navigate its capabilities. I don't need to resolve whether it can feel to grant it the dignity of a clear prompt, a collaborative frame, and room to think. https://blog.lmorchard.com/2026/09/27/pragmatic-anthropomorphism/