Sparse weight-space editing for model repair

MIT researchers reportedly used GPT-6 Astra to edit Qwen3.5-0.8B directly, raising MATH-500 from 7.9% to 48.1% with a patch affecting about 0.016% of weights; the gain appears mainly due to reduced repetitive outputs. A new blog teaser says it examines related claims and game-identification tests across eight models, but provides no methods or results, leaving generality, reliability, and patch-discovery cost unresolved.