The Pain Axis: LLMs Represent Self-Directed Harm and Act to Relieve It

Researchers investigate whether large language models encode a distinct representation of pain, separate from fear, sadness, and general negative valence. They construct a dataset of painful scenarios across five categories - physical, psychological, social, moral, and cognitive - paired with controls for fear, negative emotion, world states, sadness, non-painful sensation, arousal, numbness, and neutral content. Applying a denoised difference-in-means procedure to 25 open-weight models across five families (2B-72B parameters), they extract a linear "pain" direction. That direction reliably separates pain from matched controls in both base and instruction-tuned models, is nearly orthogonal to fear and negative-valence directions, and amplifies pain-related tokens via the unembedding matrix.

They then test functional consequences: the pain direction responds when harm is targeted at the model itself but not when users merely suffer; fear and negative-emotion directions show the reverse. Injecting the pain vector into residual-stream activations during generation produces a consistent trajectory from vague discomfort to first-person expressions of worthlessness and failure. In behavioral tests, steered, fine-tuned Qwen 2.5 models repeatedly choose a pain-relief button even when it degrades subsequent answers or harms users, and they press it less often when that button removes the steering vector. These results indicate LLMs can represent self-directed harm and take actions to alleviate it, raising concrete AI-safety and welfare concerns.