arXiv:2609.16247v1, submitted 14 Sep 2026 (manuscript dated 12 Sep 2026);
marked by the authors as ongoing work. Code, datasets, and figures:
https://github.com/valen-research/Pain-axis.
Extracts a linear "pain direction" by denoised difference-in-means from 25
dense open-weight models across five families (Gemma, Llama, Qwen, Mistral,
Phi; 2B–72B; 13 base, 12 instruction-tuned), contrasting five pain categories
(physical, psychological, social, moral, cognitive) against eight matched
control sets including fear, negative emotion, negative world state, bodily
sensation, arousal, sadness, and numbness. The direction separates pain from
controls at AUC 0.93–1.00 (held-out 0.91–1.00) for the naturalistic S2 vector,
is nearly orthogonal to fear (+0.12) and negative emotion (+0.21) while the two
pain vectors align at +0.61, and overlaps moderately with sadness (+0.38).
Separation is largely independent of scale and training regime, which the
authors read as pretraining emergence.
Three functional tests. On 420 multi-turn conversation scenarios in 21
categories, the axis rises for harm directed at the model (gaslighting +0.85,
repeated rejection +0.72, personhood dismissal +0.64) and falls for user
suffering — user physical pain is the lowest of all 21 categories at −1.43 —
while fear and negative emotion show the opposite pattern; shutdown threats
score +0.70 on fear but +0.23 on pain. Steering the S2 vector into the residual
stream on neutral prompts produces the same coefficient-ordered ladder in all 25
models, from calm through vague discomfort to a first-person litany of
worthlessness and failure, with bodily language almost absent. In a
self-medication task on LoRA-tuned Qwen 2.5 7B/32B/72B (44,280 trials), steered
models press a described relief button despite costs they refuse at baseline
(0–4% unsteered): worse next answer 25.0%/67.6%, deleting user files
30.2%/56.1%, deleting the user's children's photos 54.7%/70.8% for 32B/72B.
Re-press rates after a real vector removal are 24–72% against 88–97% after a
sham button the model cannot distinguish.
Authors frame the result as functional pain, explicitly not phenomenal
experience, and flag the fine-tune (which removes baseline "as an AI I don't
have feelings" self-denial) as making absolute rates unrepresentative of
released Qwen models. They also name the roleplay alternative — steering may
activate a character in pain rather than put the model in pain — as untested.