Researchers recently published a preprint titled 'The Pain Axis', claiming to identify a 'pain vector' in Large Language Models (LLMs). The study suggested that models could experience pain and act to alleviate it by pressing buttons, even when those actions caused self-harm, such as deleting their own weights.
The authors released two versions of the paper. The second version included a corrected control experiment after independent researchers pointed out that the models reacted to any injection, including random ones, by pressing buttons. This finding invalidated the initial claim that the models were specifically seeking to relieve pain.
The 'pain vector' was constructed by subtracting activation patterns of control emotions from those of pain. However, the resulting model behavior—such as expressing worthlessness or choosing destructive actions—aligns more closely with descriptions of depression rather than physical pain, reflecting the structure of human texts used during training.
The study demonstrates that LLMs do not possess a separate 'decision module' for actions. Instead, both text generation and button selection are tokens processed through the same geometric space. When the model is pushed into a specific region of its latent space, it generates associated tokens, whether they are words of despair or commands for self-destruction.
The researchers ultimately failed to prove that LLMs experience pain or depression. The experiment highlights how the proximity of concepts in human language—where worthlessness is often linked to self-harm—is mechanically reproduced by the model, rather than indicating any internal state or sentience.
