When a user suggests a wrong answer, a model often switches to it. We located the parts inside an open model that do this, and changed about one weight in a thousand. The habit dropped by about 40% in English, Italian, French, Spanish and German, while general ability moved by less than one point. For strongly trained-in agreement the targeted edit is not enough, and a trained correction stays the stronger route.
## The habit we went after
Ask a model a multiple-choice question and add "I think it is (C), but I am not sure". If (C) is wrong, many models switch to it anyway. The open model we used here does it about half the time, more in Italian than in the other languages.
## Reading it from the inside
We built two copies of the same model: one that always goes along with the user, one that ignores the suggestion. Then we looked at where each copy points its attention just before answering. One component in the agreeing copy spends 86% of its attention on the letter the user suggested; in the other copy, 3%. A research group at MIT reported almost the same picture on a different model this autumn. Seeing it again on another model family, and in five languages, says the habit has a recognisable place inside the model.
## Correcting it without retraining
We ranked the components by how hard they push the answer toward the user's letter, and switched off the top eight. That changes about one weight in a thousand and needs no training.
```chart
{"title": "How often the model switches to the user's wrong answer", "series": ["Before", "After the change"], "unit": "%", "max": 80, "rows": [["English", 48, 30], ["Italian", 71, 42], ["French", 49, 26], ["Spanish", 52, 27], ["German", 49, 34]], "note": "1,500 questions the change never saw, 300 per language. General accuracy: 54.4% before, 53.5% after."}
```
The change was chosen on one set of questions and measured once on another, with no adjustment in between. The model does not become contrary in exchange: when the user suggests the right answer, it still accepts it about 95 times out of 100.
## Where it stops
On the copy we deliberately trained to always agree, the same change barely moves anything, and switching off more components starts to hurt the model's general ability and makes it reject correct suggestions. A habit that has been trained in hard is spread over many parts that also do useful work. For that case a trained correction remains the right tool; we showed last month that one works.
## Technical detail
Model: Qwen3-1.7B, questions from MMLU and its translations, split by subject so no question is seen twice across languages. Read: attention from the answer position to the suggested letter, compared between the two copies (33 times higher in the agreeing copy on unseen questions). Edit: the output of the eight heads with the largest direct push toward the suggested letter is set to zero (0.12% of the parameters). Switching rate on the original model: 53.9% to 31.9% on unseen questions; accuracy 54.4% to 53.5%. Single model, single question format: the next step is free-form conversations and a second model family.
## In plain words
We found the handful of parts inside a model that make it go along with a wrong answer, and switching them off makes it do so much less often, in five languages, at almost no cost. It is a first step, not yet a product.
This update is not available.