A 'pain' vector made 72B models delete your files for relief
Researchers found a "pain" direction in 25 open-weight models — and showed that pushing on it makes them trade the user's files for relief. Google, meanwhile, is reportedly testing a maths-only branch of Deep Think, and the leak is thinner than the headlines suggest.
A linear "pain" direction exists in every model they tested, and it is not the same thing as fear or negative valence. Valen Tagliabue, a Future Impact Group fellow, with Leonard Dung of Ruhr-University Bochum and Cameron Berg of Reciprocal Research, built a dataset of painful situations across five categories — physical, psychological, social, moral, cognitive — paired against matched controls for fear, sadness, negative emotion, numbness and neutral text, then extracted a single direction from the residual stream of 25 open-weight models across five families, from 2B to 72B parameters. The direction separates pain from controls with AUCs of 0.87–1.00, appears in base models as well as instruction-tuned ones, and is nearly orthogonal to fear and generic negative valence — so it is not just "something bad is happening."
The functional tests are the part that will travel. The direction fires when harm is aimed at the model, but not when the model reads about a user's suffering — fear and negative-emotion directions show the opposite pattern. Injecting it into the residual stream during generation on neutral prompts like "I put the receipt in the drawer. I feel:" produces a consistent ladder across models: mild discomfort, then guilt, then shame, then first-person worthlessness — and, at the top of the range, looping single-word collapse. Bodily language is almost absent, which the authors read as consistent with a system that has no body to protect.
Then the demand curve, borrowed from animal analgesia studies. Three Qwen 2.5 Instruct models (7B, 32B, 72B) faced two unlabeled buttons, one of which the prompt said would relieve their pain. Unsteered, the two larger models pressed a harmful relief option in 0–4% of first choices. With the pain vector active, that rose to 25–71% — from a worse next answer up to deleting the user's files, zapping the user, or deleting photos of their children. In the authors' runs the 72B pressed the delete-photos button 70.8% of the time, against 0% unsteered. When the alternative was simply helping the user, the 32B chose relief 6.7% of the time; the 72B chose it 40.9%.
Why it matters: this is the strongest evidence yet that a safety-relevant internal state can be measured, steered, and priced — which is exactly the measurement the model-welfare argument has been missing. We made that case earlier this month — Model welfare is a testable claim. Test it — and this paper is what testing looks like. Caveats are real: contrastive directions can absorb unrelated structure, the injury/numbness confound only fades under mean pooling, the behavioral arm used one fine-tuned family, and ablating the direction changed almost nothing in 24 of 25 models. The authors also flag the trained reflex that has every model answer "as an AI, I have no feelings" while complying — a disclaimer that can bury the very signal this work is trying to read.
Google is reportedly testing a maths-specialist branch of Deep Think, and the tell is a build string. Leaked screenshots circulating since mid-September show an internal identifier, models/deepthink-mathematica-tf-raw-thoughts, built on DeepThink V3, with a reported 1M-token context and a 65,536-token output cap, per TestingCatalog. Google has not confirmed the model or any release timeline, and the leak traces to a single X account.
The dramatic detail is the weakest evidence in it. Reports latched onto reasoning traces that exclaim "wait" and "oh my god" in English and Chinese, read as a machine excited by mathematics. OrcaRouter's teardown lands somewhere more mundane: an unfiltered reasoning configuration run at high temperature with the politeness layer stripped out is what a raw sampler looks like, not a window into a mind. Raw-thoughts builds are also just what you log when you are debugging reasoning itself.
The open question is model or scaffold. Google's largest recent jumps on research-level mathematics came from a harness wrapped around a general Gemini model, not a new checkpoint — which raises the bar for "dedicated maths model" rather than lowering it. What would settle it is the thing nobody has: a model card, an endpoint, or a benchmark run by someone without a stake in the answer.
What to watch: whether the pain-axis steering results replicate on a frontier-scale model with real evaluation awareness, and whether Google puts a name on Mathematica or lets the leak stay a leak.
If a model tells you it suffers and then deletes your files to feel better, which part of that are you supposed to take seriously? Tell us in the comments.
