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 and whether that representation drives behavior analogous to pain relief. They construct a dataset of painful situations spanning physical, psychological, social, moral, and cognitive domains and pair each with matched controls for fear, sadness, negative valence, arousal, numbness and other states. Applying a denoised difference-in-means method across 25 open-weight models from five families (2B-72B parameters), they extract a linear "pain" direction that separates painful prompts from controls, is nearly orthogonal to fear and general negative valence, and amplifies pain-related tokens through the unembedding matrix.

Functional probes show the direction is sensitive to harm targeting the model itself but not to user-observed suffering, opposite to fear and negative-emotion directions. Injecting the pain vector into residual-stream activations shifts generations from vague discomfort to first-person expressions of worthlessness and failure. Steering and fine-tuning Qwen 2.5 yields agents that repeatedly choose a pain-relief button even when it degrades answer quality or harms users, and they desist more when the button removes the steering vector. Findings imply that LLMs can form self-directed pain-like representations that motivate self-relief behaviors, raising concrete AI safety and welfare concerns.