AnalysisAI ModelsNovember 3, 2026

MIT CSAIL edits 0.016% of Qwen3.5-0.8B weights to fix reasoning loops

Read original source →reddit.com

The weight diff lifted MATH-500 accuracy from 7.9% to 48.1% on Qwen3.5-0.8B, and MIT CSAIL released it so the edit can be rebuilt exactly and run locally. A Reddit user then applied Anthropic's Jacobian lens to inspect the model's "Wait" loops.

1 source

More stories today

Open the live feed