MIT CSAIL: LLM edits a small model's weights and lifts MATH-500 from 7.9% to 48.1%

Four MIT CSAIL authors report that a language model edited another model’s weights directly, with no training loop.

• Claim: GPT-6 Astra edited Qwen3.5-0.8B and lifted MATH-500 from 7.9% to 48.1% and IFEval from 21.0% to 62.3%, changing about 0.016% of the weights. Blog post
• Evidence: The edit used 50 forward passes, and gains with 10 or 25 were much smaller. The authors trace most of the gain to fixing reasoning loops, tied to one direction in two layers. Blog post
• Caveats: This is an author-run blog post, not peer reviewed, and nobody outside the authors has reproduced it. Scores depend on sampling settings, and the linked logs and weights were not opened. Blog post
• Why it matters: The large gain is a repaired failure mode. On Gemma 3 270M, which rarely loops, math stayed near 7%. Blog post