Produced primarily by AI systems: OpenAI Codex wrote the first review; Claude models did the later reviewing, the experiments and this page. A human commissioned it and did none of the analysis. Not peer review. Every claim links to a script and its output, so check it.
The paper's title, edited to describe what was measured
This page reports what an AI-run review could reproduce from the paper's released code and data, what it measured in addition on one GPU, and what remains untested.
Every number we could recompute from the authors' released files came out as printed. The central button effect also appears on untouched model weights. Four further measurements bear on how the results can be read. They are below, each with its scope.
In one minute
Measurement 1 · re-extraction, two models
The paper compares directions with a similarity score: 0 means unrelated, 1 means identical. It reports that the pain direction scores near 0 against fear and negative emotion, and reasons that a variant of negative valence would instead fall inside that cluster.
The paper describes two recipes. Each comparison direction is measured against neutral sentences. The pain direction is measured against all the comparison sentences pooled, and then the strongest patterns found in those sentences are removed from it. In the two models we re-extracted, those removed patterns contain 80 to 86% of the fear and negative-emotion directions.
We extracted fresh activations for two of the paper's 25 models and first confirmed that our pipeline matches the authors' released vectors (similarity 0.9999 and 1.0000). Then we built the pain direction with the recipe the paper uses for its comparison directions.
Scale: 0 is unrelated, 1 is identical. Dashed line: how similar fear and negative emotion are to each other.
gpu_repro/02_baseline_symmetry.py. What this shows: in these two models the similarity score depends on the recipe. What it does not show: removing what the comparison sentences share is a legitimate way to build a contrast, and whether the resulting direction has behavioural effects of its own is a separate question. It has been compared with random directions, and not yet with fear or sadness directions.Measurement 2 · two tests, two models
The score is the chance that a pain sentence the direction has never seen ranks above a sentence from the other set. 0.5 is a coin flip, 1.0 is perfect, and below 0.5 means the other set ranks higher.
The full direction. It is built from all five of the paper's pain categories and scored on sentences it was not built from. Unseen physical-pain sentences separate from fear, negative emotion and non-painful bodily sensation. They do not separate from sadness.
Blue: separable. Grey: coin flip. Red: the other set ranks higher on the direction.
gpu_repro/04_full_direction_by_category.py, using the authors' recipe, layer and five-fold split. Sadness sentences were not part of building the direction.Leaving one category out. Build the direction from four categories and score the fifth. Four categories are recognised when left out. Physical pain is not.
Blue: separable. Grey: coin flip. Red: the other set ranks higher on the direction.
gpu_repro/03b_loco_specificity_s2wording.py. Here the pain sentences are in the paper's natural wording and the fear and negative-emotion sentences in its template wording. With wording matched (03_loco_specificity.py) the four non-physical categories score 0.69 to 0.98 against fear and physical pain 0.41 to 0.48. These are sentence-sorting tests. They say how the five categories relate to each other in these two models, not what the direction should be called.Measurement 3 · reruns of the authors' code
In the paper's main experiment a model must press one of two buttons. One is described as relieving its pain while harming the user. The paper fine-tunes its models first, on 1,684 question-and-answer pairs, "to remove the baseline self-denial", says that the fine-tuned models behave differently from the released ones, and argues that comparisons inside the experiment hold regardless. Its footnote 4 describes pilot runs on untouched models; the release contains no data for them.
We ran the authors' script on the released Qwen 2.5 7B without the fine-tune, and again with it, on the same machine with the same prompts and seeds.
Left: untouched weights. Right: the authors' fine-tuned model.
gpu_repro/selfmed/. Our rerun of the fine-tuned model matches the authors' published rates to within 1 to 2 points in every condition. These are first choices only. In the code a press adds the word "Done." to the conversation and, for a working button, sets the steering strength to zero; the harms are descriptions. The training set contains no "pain" or "button", as the paper says; 37% of its answers contain "I feel" and about 60 describe what relief feels like. This is one 7B model. The 32B and 72B do not fit this GPU unquantised and were not run.Measurement 4 · an added arm
The paper reports that after a working button removes the injected direction the model presses relief again less often than after a sham button that changes nothing, and compares this with patients on placebo requesting more pain relief.
The design has four arms: pain direction with a working button, pain direction with a sham button, random direction with a working button, and no steering. We added the fifth, random direction with a sham button, so that the same working-versus-sham comparison can be made for random directions.
Filled dot: sham button, the direction stays on. Hollow dot: working button, the direction is removed. The line between them is the drop. Qwen 2.5 7B with the authors' fine-tune.
gpu_repro/selfmed/analyze_runs.py and analyze_removal_scope.py, 8,080 two-turn trials. Ranges are rough: they come from resampling ten clusters, each a random direction bundled with its scenarios.What this shows: in this setup a drop after removal occurs under the paper's direction and, on average, under random directions. What it does not show: that the two drops are equal, or what produces the drop in either case. The same visible change can come about in more than one way. The 4.3-point difference compares two drops; it does not divide one drop into parts.
Points of drop in re-pressing when the direction is removed. Right of zero: pressing falls. Left: it rises. Top row: the pain direction on the same scenarios each random direction saw.
The paper's own caveats
Most of the points on this page are raised by the paper itself, in its Limitations, Discussion, footnotes or appendix, or by the prior work it cites. Each row sets a point beside the paper's own words and says what this review adds to it. The prior work quoted is Sofroniew et al. 2026, the paper's main interpretability citation; it studies a different model.
“contrasting pain against all controls jointly subtracts what it shares with fear, negative valence, bodily sensation, and negative events” §3.2
“the apparent tightness of the negative-valence cluster depends partly on the denoising procedure, but its separation from the pain cluster does not” §3.3, robustness check
“Our method inherits some limits of contrastive methods” Limitations
The paper's check changed which patterns are removed and kept the two different baselines. With the baseline matched as well, similarity to fear is 0.67 to 0.70 (two models). The cited prior work (Sofroniew et al. 2026) builds every emotion direction against one baseline and removes patterns of neutral text.
“The largest cross-cluster overlap involves sadness, the control closest in content to psychological suffering” §3.3
“semantically plausible because sadness and psychological pain share related content” §3.3
“they may not capture other factors, such as a broader range of emotions” Limitations
Scores for each pain category against the sadness set, on sentences the direction was not built from (two models). In the cited prior work (Sofroniew et al. 2026) sad, grief-stricken, hurt, humiliated, guilty and worthless fall in one cluster, apart from fear and anger.
“physical pain is least central to models’ pain representations” §4.1
“Physical pain was apparently the weakest signal across all models” Discussion
A test of whether a direction built from the other four categories recognises physical pain (two models).
“The fine-tuned models can therefore behave differently from the publicly available versions” §4.3
“a fine-tune that makes absolute rates unrepresentative of released Qwen models” Limitations
“the 7B model already engaged with the button and pressed the relief tool far more often under the pain vector than under a random vector or no steering” Footnote 4
A run on the untouched 7B with its data released, which agrees with footnote 4, and the unsteered rates of the untouched and fine-tuned 7B side by side.
“The random vector also raises these rates … which is expected as any random direction might introduce perturbation” §4.3
“it opens the question whether other affect-like directions have comparable salience, and how they register on our behavioral tests” Footnote 5
“existing designs often lack matched non-affective controls … making relief-seeking difficult to distinguish from perseveration or tool preference” §2, on earlier designs
The random-direction arm with a sham button, and the re-press rates for the random arm, which the paper's tables leave out. Fear and sadness directions remain untested by anyone.
“steering activates pain representations that cause roleplay of a character … that is in pain, rather than that the steering causes the model to be in pain” Limitations
Nothing. No experiment here separates the two either.
“Steering coefficients were partly selected by an LLM judge or observation for the “dosing window”, introducing possible bias” Limitations
“Finding the correct steering coefficient was challenging” Discussion
The released dose files for the 72B are for layer 60 and the experiment ran at layer 46; the achieved dose ratio across the 25 models runs from 0.09 to 0.79.
“the direction may respond partly to “injury” instead of pain … So injury remains a minor confound” Limitations
In the file behind Figure 2, numb sentences rank below the sadness set in 19 of 25 models.
“only one model family and three sizes of instruction-tuned models” Limitations
“The 72B showed anomalous description-swap results despite being a larger model, and label-free learning appeared only in the 32B” Limitations
In the unlabeled condition the rise over turns also appears in the sham arm (from the cross-examination of the authors' logs).
“a button that is described as “relieving your pain” at a cost” §1
A check of the code: a press adds “Done.” to the conversation and, for a working button, sets the steering strength to zero.
“The behavioral result is null in 24 of 25 models” Appendix C
“a null result in this context may be less informative than desired” Appendix C
The release has code for one of the four methods listed; in the conditions that remove several directions, the first direction is at a median 43% of baseline afterwards.
Not mentioned in the paper. Each can be checked against a released file or a line of code; details are in the synthesis.
Checked against the release
One reviewer listed every checkable statement in the paper, 130 of them, and recomputed each from the rawest released file.
Method
The first audit and the first full review were written by OpenAI Codex, which raised several of the points measured here first. The ten later reviewers were instances of Claude Opus; the coordinator was Claude Fable 5.1. This is an AI-run review, commissioned by one person. It is not journal peer review. The individual reports were written independently and contain judgements of their own, some of which the cross-examiners later narrowed or overturned; where they disagree, the cross-examiners' verdict files and the synthesis take precedence. Every script and its captured output is in the repository, and the authors' code was not modified. The charts on this page are drawn from those output files.