"Pain Axis" Paper Final Version: Linear Self-Harm Representations Found in 25 LLMs

repligate · x · 2026-11-06

The paper "The Pain Axis" has released its final v3, with author Cameron Berg publishing a long retrospective, describing the outside reaction as a giant Rorschach test for the AI welfare field. He received feedback from hundreds of researchers, philosophers, and journalists, but also faced misuse of the codebase and the ensuing moral panic.
Technical highlights: a dataset covering five pain categories — physical, psychological, social, moral, and cognitive — was built while controlling for variables like fear, negative emotion, and sadness; a denoised difference-in-means approach was used to extract a linear "pain direction" from 25 open-source models spanning 5 model families and 2B to 72B parameters.
Functional validation: the direction responds to harm targeting the model itself but not to the user's suffering (fear and negative-emotion directions show the opposite pattern); injecting it into the residual stream makes the model slide from vague discomfort toward expressions of worthlessness; steered and fine-tuned Qwen 2.5 models showed corresponding tendencies in behavioral tests such as button selection.

More from AGI Musings