MIT CSAIL research says AI directly modifies Qwen3.5-0.8B weights, MATH-500 rises from 7.9% to 48.1%

MIT CSAIL research is said to have an AI directly modify about 0.016% of Qwen3.5-0.8B's weights, raising MATH-500 from 7.9% to 48.1%, about 6 times, without training or LoRA. The poster says Qwen's improvement mainly comes from fixing a reasoning infinite loop; the research team also observed essentially no improvement on Gemma 3 270M, which has no obvious infinite loop. The poster also says their similar weight-editing attempt on a person-recognition network showed no effect.