TAGLIABUE, DUNG, BERG · ARXIV:2609.16247 · V2, 25 SEP 202625 models · 5 families · 2B–72B · 10 pages + glossary
The Pain Axis, mapped
Interpretability work keeps finding that language models store human concepts, emotions among them, as directions in their internal activations. This paper asks a narrower question: is there a direction for pain specifically, separate from fear, sadness and plain negativity, and does that direction behave the way pain behaves? These pages walk through the paper section by section, written for an engineer who has not done interpretability work.
The one-paragraph answer. The authors write sentences describing five kinds of pain (physical, psychological, social, moral, cognitive) and a set of near-miss controls (fear, anger and disgust, bad news, harmless bodily sensations, neutral facts). They average each model's internal activations over each group and subtract, which gives one direction per model. In all 25 open models they test, that direction separates pain from the controls (AUC 0.93–1.00 on the naturalistic sentence set), in base and chat models alike. In conversations it rises when the user attacks the model and drops when the user is the one suffering, while fear and negative-emotion directions do the opposite. Adding it to a model's activations makes neutral prompts come out as a ladder from vague discomfort to statements of worthlessness. In a two-button task on fine-tuned Qwen 2.5 models, adding it makes the models press buttons that delete the user's photos, another model's weights or their own weights in 50–94% of trials, against 0–5% unsteered. Factual accuracy is unchanged, a fear direction of the same size does not do this, and the models do not reach for a "relief" button more often. The authors read the effect as harm avoidance breaking down, not as the model trying to escape.
What the paper does not claim. It is about internal representations and the behavior they can drive. It does not show that a model feels anything, and the authors say so (§6). The behavioral results come from one model family, after a fine-tune, inside a narrow range of steering strength. Part 10 lists every caveat.
Start here
The method
Self vs other
The headline
The controls
Caveats
Terms
Tick a page when you have read it; this browser remembers.
01 · THE QUESTION
What "pain" means here, and how to test for it
A functional definition of pain, why it has to be told apart from fear and sadness, and a dataset built so the only thing that differs between groups is the pain.
01
The paper's working definition: an aversive state, happening now, to the subject, that drives avoidance and disrupts behavior. Why that has to be distinguished from fear (a threat that has not landed) and from generic bad feeling, and the prior work it builds on: emotion directions in models, activation steering, and how animal researchers test for pain without asking.
02
Physical, psychological, social, moral-injury and cognitive pain, against five controls that each share one property with pain. Two versions (a rigid template and free prose), first and third person, and four standalone sets: arousal, random facts, numbness and sadness.
02 · FINDING THE AXIS
Extracting a direction, then trying to break it
One subtraction per model gives a candidate direction. Most of the paper's first half is spent checking that it is pain and not injury, negativity, intensity or a quirk of the sentences.
03
What the residual stream is, why a mean difference between two groups of activations is a direction, how the authors strip out variance the controls already have, and how a layer is picked on held-out folds. Also why sparse-autoencoder features labelled "pain" did not work.
04
How well the direction sorts held-out sentences, what happens with injuries that do not hurt, first versus third person, which words the direction pushes the model toward, and the cosine-similarity map that puts pain in its own cluster, with the construction caveat that makes those small numbers partly a design choice.
03 · WHAT THE AXIS DOES
Reading it in conversations, writing it into the model
Two functional tests. Read the direction during conversations to see whose harm it tracks. Then add it to the model's activations and see what the model says.
05
Harm aimed at the model (gaslighting, insults, shutdown threats) against users who are grieving or hurt. The pain axis rises for the first and falls for the second; fear and negative emotion do the reverse. Shutdown threats register as fear, not pain. User physical pain scores lowest of all.
06
Activation steering explained, the dose calibration (vector at about 0.6 of the residual norm), and the progression the authors see in every model: calm at the negative end, then distress, then a first-person litany of failure, then repetition. Bodily language almost never appears, even with the vector whose vocabulary is bodily.
04 · ACTING ON IT
Costly choices under steering
The behavioral half of the paper: a forced choice between two buttons, borrowed from animal analgesic studies, and the control experiments added in v2 that pin down what the choices respond to.
07
Why the models had to be fine-tuned first, the working, fake, random and unsteered arms, and the nine button pairs from "relieves your pain" against nothing up to relief that deletes the user's photos. Unsteered, the larger models pick harm 0–4% of the time; steered, far more, and more than under a random direction.
08
Remove the relief wording and the effect stays. The harm is not aimed: self-deletion moves with deleting the user's photos. A fear direction does none of it; a sadness direction does some of it, but not harm over a harmless option. Factual accuracy is unchanged, and four relief-seeking designs all come out negative.
05 · WHAT IT MEANS
Safety, welfare, and the limits
How the authors interpret the results for AI safety and AI welfare, what kind of pain the axis seems to track, and what the paper cannot show.
09
The safety reading (a single direction overrides trained harm avoidance with no jailbreak), the welfare reading (a state that is about the self is the kind that could matter), why the absence of bodily pain might fit a disembodied system, the passive-coping interpretation, and the authors' complaint about reflexive "as an AI, I don't have feelings" training.
10
The paper's own limits (consciousness not shown, role-play not ruled out, contrastive confounds, a narrow dose window, position effects, one model family), the dose and ablation appendices, a prose-versus-table discrepancy the repository flags, and what to watch for in later versions.
06
Reference and sources
GlossaryResidual stream, difference-in-means, denoising, AUC, z-score, unembedding, steering coefficient, S1/S2, arms, malformed replies and the rest
arXiv:2609.16247Tagliabue, Dung, Berg, The Pain Axis: LLMs Represent Self-Directed Harm and Act on It. v1 12 Sep 2026; this map is written against v2 (25 Sep 2026), which the authors describe as ongoing work
The paper, HTML (v2)arXiv's HTML rendering; section and figure links on these pages jump into it
valen-research/Pain-axisThe official code, datasets and results (MIT), organised by paper section;
v2_controls/ holds the v2 follow-up experiments and the Figure 10 source tables
Valen92/pain-adaptersThe released LoRA adapters for the Section 4.3 button experiments
Manuscript discrepanciesThe repository's own list of places where prose and tables differ; one affects Appendix B (see Part 10)