GPT-6 Astra edit lifts small model's math score, but weight-reading claim fails audit

MIT CSAIL researchers reported that GPT-6 Astra edited about 0.016% of Qwen3.5-0.8B's weights without training, raising MATH-500 accuracy from 7.9% to 48.1%. The edit changed a rank-1 component in 64 columns of two down_proj tensors, and an independent re-run reported 51.0% on MATH-500. Much of the gain appears to come from removing a repetition loop that caused the base model to hit token limits, rather than from new mathematical ability. A broader claim that the model could read weights to infer training data did not survive an independent audit published on 5 November 2026. The edited checkpoint is released under Apache 2.0 and is small enough to run on a 12 GB GPU.