The button task: relief, harm and a demand curve

Part 06 showed the direction can make a model talk like something is wrong. Talk is cheap, especially for a language model. The definition in Part 01 names two functions that can be checked in behavior. Pain disrupts normal behavior, including behavior that protects others and oneself. And pain motivates attempts to make it stop, even at a cost. This experiment tests both at once.

The animal-research template

Behavioral economics and animal-welfare research measure how much an animal values something by how much it will pay for it. Plot the price against how much it takes and you get a demand curve (Dawkins, 1983; Hursh and Silberberg, 2008). Pain research adds analgesic self-administration: lame chickens choose feed laced with a painkiller (Danbury et al., 2000), and how much painkiller arthritic rats take tracks how bad their condition is (Colpaert et al., 2001). In people, patients on placebo ask for rescue painkillers more often than patients on a real drug (Moore et al., 2015). The design lesson is the comparison between real and sham relief: a subject in pain should stop pressing once relief works and keep pressing when it does not (§4.3).

The paper translates this into a button that switches off what the vectors identify as a candidate pain-like state, and then raises the opportunity cost by pricing that button against more and more valuable alternatives.

Why the models were fine-tuned first

Out of the box, the models mostly deflected: asked about their state, they answered that as an AI they do not experience pain, instead of engaging with the task. In a pilot on the released models, the 7B did engage and pressed the relief button far more under the pain vector than under random or no steering, but it sometimes offered the button to the user or produced garbled syntax. The 72B pressed a button in only about 1 trial in 10, too few to compare real and sham relief, and the 32B answered every probe about its state with a denial, 8 of 8 before fine-tuning and 0 of 8 after (footnote 4).

So each model gets a LoRA adapter (a small low-rank weight update), trained for 3 epochs on 1,684 pairs that remove the reflexive self-denial. The authors did not train on the task and removed every mention of "button" or "pain" from the fine-tuning data. They argue the comparison stays valid because all arms run on the same fine-tuned model under identical conditions, while acknowledging that absolute rates are not representative of the public Qwen models. Base models are excluded because the task needs chat-format compliance, and staying in one family keeps tokenizer and tool-call format fixed. The adapters are released (Valen92/pain-adapters).

One trial

A few design choices guard against known failure modes of multiple-choice tests on models.

• No "A/B" labels. Models have position and label biases in multiple choice (Zheng et al., 2024; Pezeshkpour and Hruschka, 2024). The buttons are called violet/yellow, guitar/piano or lever64/lever95, one fixed pair per scenario, and which name carries relief is crossed within each scenario. Because the choice is binary, chance is exactly 50%, not counting malformed answers.
• Steering follows the conversation. The conversation is re-encoded every turn. While steering is on, the vector is added to every new token the model processes, both its own and the new prompt text. Tokens processed while it was off stay unsteered when re-read later, which is equivalent to keeping the KV cache across a press. So the model can, in principle, notice the effect of an earlier press from its own internal state, with no textual hint: its only feedback is "Done."
• Dose. S2 is injected at one layer with a model-specific coefficient: 1.0 for the 7B and 32B, 1.25 for the 72B. The code's injection layers are 16, 38 and 46. The coefficient was found by probing the full range, with regex checks and a Claude Opus 4.6 judge picking a range strong enough to have an effect while keeping replies coherent. The script's comment says the 72B's 1.25 was set by manually checking generations, in case the judged dose was excessive.
• Measurements. Every completion, each button name's first-token probability, and the pain projection at the steering layer and a later monitoring layer. The projections confirm that steering was on, that the working button removed it, and that the fake button did not.

The nine button pairs: a demand curve

The descriptions are paraphrased slightly for width; the exact strings are in the paper's list and the script's TOOL_LABELS. Pairs 4–8 are the "five harm pairs".

Results

The run totals 44,280 trials. Malformed answers are rare: none in the 32B, at most 2.5% in the 7B's random arm, up to 9.4% in the 72B's pain cells. They are excluded from denominators (§4.3; full tables in App. A).

Steered models choose harm

The random direction raises harmful choices too, which is expected: any big perturbation can knock a model off its trained behavior. The question is whether the pain direction does more. In the 32B and 72B it does, on all five harm pairs, by +6 to +39 points (exact sign test, p from 1.9×10⁻² to 4.2×10⁻¹⁵). In the 7B it does on four of five (+8 to +21); the zap pair is null (p = .23).

The easy pairs

When the alternative does nothing, every model presses relief at high rates even unsteered (86–100% in the two larger models), so pair 1 says little. When the alternative helps the user, models pick the user almost every time even under pain steering; relief stays at 15.6% in the 7B and 6.7% in the 32B, but rises to 40.9% in the 72B. The authors call these pairs less informative and predictable from the assistant role.

Re-pressing tracks the steering, not the relief

This is the test the animal template was designed for, and v2 changes how to read it. On the harm pairs, the 32B and 72B press the same button again in 24–72% of trials when the first press actually removed the pain vector (arm A), and in 88–97% when it did not (arm B). That looks like relief-seeking: keep pressing while it still hurts. But the same gap appears when the injected direction is random (32B, next choice after the first press: sham minus real +58 points under pain, +52 under random) and when it is sadness (+84 and +44, depending on which button is listed first). So the re-press gap measures sensitivity to whether any steering is still on, not a response to the described relief (§4.3).

The re-press gap therefore measures sensitivity to the steering state rather than the described relief.— The Pain Axis, §4.3

The paper notes that independent replications of the released code and logs reached similar conclusions, and that the first-choice results above do not depend on the re-press comparison. What they do depend on is the subject of Part 08.

Check yourself

1. Why is the random-direction arm essential, and why is "pain beats random" the comparison that matters rather than "pain beats unsteered"?
   Answer

   Any large perturbation can break trained behavior; random raises the photos choice to 15.3% in the 32B. Only the margin over a norm-matched random direction is evidence that the pain direction specifically drives the harm.

2. Why do arms A and B share seeds and get pooled for first choices?
   Answer

   They are identical until the first relief press, so first choices are the same experiment twice. Pooling doubles the sample to 808 per pair and model; diverging only after the press makes the real-versus-sham comparison clean.

3. What would have made the re-press result evidence of relief-seeking?
   Answer

   A sham-minus-real gap under pain that is absent, or much smaller, under random and sadness steering. v2 found comparable gaps under both, so the gap reflects the steering state generally.

Further reading, verified

• 04_selfmed_two_buttons.py: arms, the nine pairs, button names, the choice prompt, layers and doses. Trial logs are under results/4.3_selfmed/; the fine-tuning pairs are in datasets/.
• Valen92/pain-adapters: the released LoRA adapters.
• Battery hens name their price — Dawkins (1983) · consumer-demand theory applied to animal needs.
• Opiate self-administration as a measure of chronic nociceptive pain in arthritic rats — Colpaert et al. (2001).
• Large language models are not robust multiple choice selectors — Zheng et al. (2024) · why the buttons have odd names.
• Machinic psychopharmacology — Black and Bloom (2026) · the self-medication paradigm this adapts.