MIT CSAIL researchers say GPT-6 Astra fixed Qwen3.5-0.8B by editing 0.016% of its weights, without training
In a blog post the authors call very early results, GPT-6 Astra read raw weights, traced Qwen's reasoning loops and patched them by hand, lifting MATH-500 from 7.9% to 48.1% and IFEval 3x; it is not peer-reviewed.
Why it matters
Changing what a neural network does has normally required gradient-based training or a purpose-built tool. This post reports that a general-purpose frontier model, prompted in plain language, can read raw weights, identify what a network was trained to do, diagnose a failure that Qwen's own model card warns about and patch it directly. The authors stress the results are very early and published as a blog post, not a paper, but they released full traces on GitHub. Watch for independent replications and the benchmark and harness the team says it is building.
Key facts
- GPT-6 Astra edited 0.016% of Qwen3.5-0.8B's weights
- MATH-500 rose from 7.9% to 48.1% with the same sampling settings
- IFEval instruction following rose 3x
- No gradients or training loops were used
- Astra named the game for all four PPO-trained networks from their weights
- Lead author Rikiya Takehi lists MIT CSAIL