Researchers edit Qwen’s weights to reduce reasoning loops

MIT CSAIL researchers report using GPT-6 Astra to inspect and directly edit Qwen3.5-0.8B’s weights, reducing repetitive reasoning loops. With 50 allowed feedback passes, MATH-500 rose from 7.9% to 48.1% under their evaluation settings. This targeted repair helped answers finish; a separate attempt to improve a model without the looping problem produced no established gain.

Rikiya Takehi and collaborators · MIT CSAIL · Nov 2, 2026

Why it matters

The experiment separates removing a specific failure mode from adding general reasoning ability.

More context

Astra used a Python environment to inspect tensors and practice problems, without gradient-based training. The reported edit changed 131,072 weights, or 0.016% of the model. Held-out evaluations used the same sampling settings for the original and edited versions.

Keep in mind

These are early author-reported experiments, not an independently replicated repair method. Feedback budgets and model-specific failures matter; larger models and unrelated defects may behave differently.

Original read