Does a Weight Edit Teach Qwen3.5-0.8B Math, or Stop It Looping? | hedwigai

An analysis of an AI weight-editing experiment on the Qwen3.5-0.8B model reveals that modifying a small fraction of its parameters does not teach the model math, but rather acts as a general brake on repetition and thinking loops. Researchers found that the edit removes a specific shared direction in the residual stream when neurons fire during repeated text. Consequently, the model's performance gains primarily stem from escaping loops rather than acquiring new reasoning capabilities. Open article →