Astra Adjusts Qwen's Weights So the Model Can Finish Its Reasoning to the End

At a glance

MIT researchers asked OpenAI's model GPT-6 Astra to directly adjust the internal numbers of Qwen3.5-0.8B, a small AI model. By reducing the model's tendency to loop around the same point, the share of problems it finished solving rose sharply.

Key points

1. 1

   Astra changed only about 0.016% of the model's 800 million weights.

2. 2

   After the change, the share of answers cut off by the token limit fell from 93% to 33%.

3. 3

   The MATH-500 math score rose from 7.9% to 48.1%.

4. 4

   The authors explain that the math ability was not newly created but that obstacles hiding the model's existing ability were removed.

Why it matters

It shows that the problem of AI failing to reach a final answer may be solved by directly editing the model's internal workings.