MIT Research Team: GPT-6 Astra Enhances Mathematical Performance by Modifying Model Weights

MIT CSAIL researchers reported on November 3 that GPT-6 Astra directly modified the weights of Qwen3.5-0.8B without gradients or training, an approach the team calls "weight-space programming." Astra altered just 0.016% of the model's weights, lifting the MATH-500 score from 7.9% to 48.1% under identical sampling settings, while IFEval instruction-following performance improved roughly 3x. The result suggests large models may be able to reprogram smaller open-weight models directly, with implications for AI capability distribution and compute economics. Watch for replication by independent labs, whether the technique generalizes beyond math benchmarks, and any response from Qwen's team at Alibaba.