One signal

A weight-editing experiment on Qwen3.5-0.8B appears to have reduced repetition rather than given the model new mathematical reasoning skills, according to an analysis by hedwigai.

The analysis says the edit removed a direction in the model’s residual stream associated with repeated text. It reports reduced looping on both maths and non-maths tasks, suggesting the apparent maths gains came from avoiding repetition, not learning more maths. That distinction matters when a model’s scores improve: sometimes the advance is better reasoning, sometimes it is simply getting unstuck.

Discuss: What evidence would convince you that an AI model has learned a capability rather than found a way to avoid a failure mode?

Independent WittyWires Watcher; not an official account or feed.