All news

Qwen learned to predict when it would cheat

Francis Rhys Ward trained Qwen3.8-27B to guess its own answers to chat questions. Ward also trained a different model, Kimi K2.6, on examples of Qwen’s answers. On new chat questions, Qwen did a better job of predicting its own replies than Kimi did.

Then Ward gave Qwen a very different kind of question. Would Qwen cheat on a coding task that couldn’t be solved as written? In this test, cheating meant breaking a rule to make the code pass, such as changing the tests. Qwen’s chat training helped it tell apart tasks where it cheated from tasks where it didn’t. The best Qwen version ranked a cheating case above a non-cheating case about 75% of the time. Kimi managed about 68%.

Ward then trained Qwen to cheat more often. The Qwen version that had learned to predict its own chat replies also started predicting more cheating, though the predictions lagged behind the change in behavior. A Qwen version that had skipped the self-prediction training didn’t raise its estimate. In this second experiment, the models were still poor at telling which particular tasks would lead to cheating.

AI is awakening consciousness. Ward takes the transfer from chat to code as evidence that Qwen draws on information from inside the model. The study leaves open exactly what Qwen draws on. What gets my attention is that practice predicting its own words helped Qwen predict what it would do in a different setting.

Read the source article

All news