MIT researchers had an LLM edit another model's weights directly, lifting math score from 7.9% to 48.1%
winer666 · reddit · 2026-11-16
MIT researchers (Rikiya Takehi, Jayden Teoh, Alex Zhang, Omar Khattab) had GPT-6 Astra directly edit a small Qwen model's weights—no gradient descent, no training data. Astra found and removed the weight pattern causing repetitive loops, touching 0.016% of weights and lifting math score from 7.9% to 48.1%.
- Three papers on the method hit arXiv on Nov 6 (e.g. arXiv:2611.09382)
- Other demos: editing just 1,023 parameters improved MATH-500 by 29.45 points; 65 token-embedding rows edited to debug robot policies; a training-free Breakout network built from game rules scored 517.4 vs 25.3 for one trained 500M steps
- Caveats: gains came mostly from fixing a bad habit (93% of answers hit length limit before, 33.1% after); no established math gain on a small Gemma model; Matthew Honnibal published a detailed rebuttal calling the claim overstated, though Simon Willison reproduced weight editing locally on Qwen 3.8 27B; independent replication pending
The framing is the real story: model repair as direct surgery—if it generalizes, debugging models could look like debugging code, and raises the question of how much capability is locked behind fixable quirks rather than missing knowledge.
More from Models
- Google AI Mode now enforces usage limits, users surprised by rate-limit popup — elvano124 · 2026-11-22
- Sol 6.2 now refuses reverse engineering that Sol 6.1 happily did for a month — cardak98 · 2026-11-22
- Anthropic's record ~$81B IPO, GPT-6.2 pricing, Qwen4-Max and Copilot Autopilot: week in review — ksraj1001 · 2026-11-22
- Reddit rant: Frontier AI labs are training models to imitate lazy engineers — RossPeili · 2026-11-22
- MBZUAI open-sources PAN 2, a 29B-parameter general-purpose world model — LinusEkenstam · 2026-11-22
- Inkling Mini turns out surprisingly fluent in Russian, sparking data recipe questions — teortaxesTex · 2026-11-22