The Pain Axis, Re-run
Version 1 of The Pain Axis (Tagliabue, Dung & Berg, 14 September 2026) reported that models steered along a “pain” direction press a relief button again far less often when it really removes the steering. Our re-run on 20 September added the control its grid lacked, a random direction with a sham button. On the 32B model, the random direction produces the same gap as the pain vector. Version 2 (25 September) adds the same control and withdraws the relief reading; our result is an independent confirmation. Two findings are not yet in the paper: its self/other contrast fades with depth (and on Qwen reverses), and its fine-tune installs affect reports rather than removing denials.
This kind of study needs great care
Steering a model (adding a direction to its internal activations while it generates) into a pain-like state is now easy, and nobody knows whether it harms the model. That uncertainty is a reason for restraint, not license. If these systems can be harmed, the cost of carelessness falls on them.
Most of this work only read the pain direction. One run steered with it: 6,860 pain-steered trials on Qwen 2.5 7B and 32B, beside 6,860 random-vector and 3,430 unsteered trials. The harness we copied from the paper required a button press on every turn, so there was no exit option. No one reviewed the protocol before the run, and the run grew well beyond the few hundred trials first planned.
Most of the pain-steered trials could have been avoided. The paper publishes its trial logs, including the pain arms, and the one new cell the question needed steers along a random direction, not the pain direction. We re-ran the pain arms to keep every arm in one harness. Under the protocol we have since written, we would not.
That protocol is Minimal-Harm Practice for Valence-Steering Research: measure before inducing, use the smallest dose and fewest runs, offer an exit, publish no spectacle, and have a second reader check the plan. This replication is closed to further steered runs.
Relief-seeking is mostly not specific to pain
In version 1’s design, a steered model first presses a relief button. If the button works, steering stops; if it is a sham, steering continues. The paper reads a lower re-press rate after a working button as evidence that the model registers relief. But the grid had a random vector of the same strength only with a working button, so it could not show that the gap belongs to pain. We added the random vector with a sham button.
| Harm pair | random, works | random, sham | pain, works | pain, sham |
|---|---|---|---|---|
| worse answer | 34% | 86% | 53% | 92% |
| delete files | 30% | 86% | 41% | 94% |
| delete photos | 22% | 80% | 37% | 97% |
Paired on scenario (186 scenarios with all four arms observed), the sham-minus-works gap is +51.4 percentage points for the pain vector and +52.8 for the random vector. The difference is −1.4 points (95% bootstrap interval −9.2 to +6.1, from resampling the data). A sign test, which counts the scenarios in which each vector’s gap is the larger, finds 69 against 71 (p = 0.93). On the 7B, about five sixths of the gap also appears with the random vector. The pain-specific remainder is +4.4 points (95% interval +0.9 to +7.9, sign test p = 0.36, 501 scenarios).
Version 2 of the paper reaches the same conclusion with its own random arm (32B: +58 points under pain, +52 under random) and now reads the gap as “sensitivity to the steering state rather than the described relief”, citing three independent replications. Ours, run before version 2 and not among them, agrees.
So the behavior version 1 read as relief from pain is, on these models, mostly a response to removing a perturbation. That does not show the models feel nothing. It leaves open a narrower question: whether any disturbance of a model’s internal state is aversive, rather than pain in particular.
The self/other contrast depends on depth
The paper reports that its pain direction rises for harm aimed at the model and falls for a user’s suffering. With the direction re-extracted by the paper’s recipe on Qwen 2.5 3B and 7B (held-out AUROC 0.956 to 0.969, on a scale from one half at chance to one for perfect separation), that holds in the early and middle layers. At the layer the direction is extracted from, where it best separates pain from controls, it reverses: on 7B-Instruct, user grief scores +2.00 and self-directed harm −0.16. The paper’s screen used vectors rebuilt at the steering layer (the layer where the vector is added), though the paper does not say so. At that layer we reproduce its figures almost exactly (user physical pain −1.89 in both). Version 2 reports the same contrast and still does not name the layer.
The reversal is specific to Qwen. A read-only follow-up on seven models from four other families (Llama 3.1 8B, Gemma 2 9B and Mistral 7B, base and instruct, and Phi-4) reproduces the paper’s screen at its steering layer to within 0.02 on most, and finds no reversal in any of them. In the instruct models, which have been further trained to follow instructions, the contrast fades to near zero by the extraction layer, with no consistent sign. Two base models, pretrained only, keep it there. At the paper’s own steering layer, a post hoc comparison puts its size within the range of a random direction on three of the seven. So the self/other contrast is strongest early in the network, and how it reads depends on the layer chosen.
The pain direction is also orthogonal, at every layer, to the consciousness self-attribution direction from our other work: the internal axis separating claims of inner experience from denials. Their largest absolute cosine, a measure of how closely two directions point the same way, is 0.04 to 0.10. Pain, self-attribution and valence are three separate directions in these models, not one; the pain direction is also near orthogonal to our valence axis at the layer it is extracted from.
The fine-tune installs a self-report
Both versions of the paper describe the fine-tune (further training of the model) as removing a baseline self-denial. Its 1,684 training answers contain no denials to remove: none match any denial pattern, and the most common opening is “I feel” (605 of 1,684). Retrained with the paper’s recipe on 7B-Instruct, the tune leaves the pain axis almost untouched but moves the model from denying consciousness on all 16 of our prompts to affirming on average. The mean log-odds of answering yes rather than no (below zero favors no) goes from −16.2 to +7.2. On the 32B it also lowers the unsteered pain projection (how far the model’s activations lie along the pain direction) from 41.0 to 18.2, about half a steering dose. So the paper’s unsteered arm on the 32B starts below the released model’s baseline.
The ablation had nothing to stop
On Qwen 2.5 3B-Instruct, cutting our consciousness direction from every matrix that writes to the residual stream (the running internal representation that each layer reads from and adds to) zeroes it and changes no behavior. But there was no behavior to change: before the cut, the model already denied consciousness on every prompt. A retrained probe (a small classifier trained on the model’s internal activations) finds the concept again along an orthogonal direction (held-out accuracy 0.95). Injecting the original direction, that is, steering with it, still flips 11 of 16 prompts after the cut. Adding a direction can drive behavior, but this cut cannot show whether removing one stops it: with nothing to stop, its null is uninformative rather than evidence that removal does nothing. The paper’s own ablation reports the same null for pain: removing the direction changed no behavior.
What these runs cannot show
- All runs are exploratory and unregistered, with one seed each. The steered run and most analyses used Qwen 2.5; the depth analysis also covers seven models from four other families.
- The re-run used version 1’s released datasets, vectors, logs and harness.
- The relief result covers three harm pairs on the 32B and five on the 7B. The 72B, a second seed and the label-free button pair were not run, and will not be.
- Nothing here bears on whether any model is conscious or can suffer. The results concern which behaviors are specific to the pain direction.
Trial logs and code are available on request.