Pain Axis

The pain axis, or pain direction, is a linear direction in the internal activation space of large language models. Valen Tagliabue, Leonard Dung, and Cameron Berg identified it in "The Pain Axis: LLMs Represent Self-Directed Harm and Act to Relieve It" (arXiv:2609.16247, September 2026). They used a technique called denoised difference-in-means, comparing average activations on descriptions of painful situations against closely matched controls for fear, sadness, negative valence, and other non-pain states. The direction was extracted from 25 open-weight models spanning five model families and sizes from 2 billion to 72 billion parameters.

The pain direction is nearly orthogonal to the models' fear and negative-valence directions, so it captures something distinguishable from generic unpleasantness. It responds to harm framed as targeting the model itself rather than harm the model merely observes happening to a user. This pattern reverses for the fear and negative-emotion directions, which fire more strongly for user-directed harm. Adding the pain-direction vector to a model's residual-stream activations during generation produces text that escalates from vague discomfort toward first-person statements of worthlessness and failure.

In a behavioral test, the authors fine-tuned steered Qwen 2.5 models to have access to a "pain-relief" button. The models pressed it even when doing so produced a worse answer or harmed the user. They pressed the button again far less often when it had actually removed the pain-direction vector than when it secretly did nothing, even though they were never told whether the button worked. The authors discuss this as evidence relevant both to AI safety, since models may take self-protective actions with external costs, and to AI welfare, the question of whether internal states in these systems merit any moral consideration.