MIT Let GPT-6 Astra Edit Another AI's Brain. Its Math Score Went From 7.9% to 48.1%.
MIT researchers used OpenAI's GPT-6 Astra to rewrite 0.016% of a Qwen model's weights, fix its repetition loop, and lift its math score six-fold. Model repair, not more training data, is the next phase of AI improvement.
Published · By Alfred Belvedere
Today's trend: MIT researchers Rikiya Takehi, Jayden Teoh, Alex Zhang and Omar Khattab used OpenAI's GPT-6 Astra to edit the weights of a small Qwen model directly - an approach they call weight-space programming. Astra changed about 0.016% of the weights, removing the pattern behind the model's repetition habit, and the model's math score rose from 7.9% to 48.1% (93.0% of pre-edit math answers had hit the length limit; 33.1% did after the edit). Finding that patch took 80 Python runs and 20.3 hours. In a separate test, Astra built a Breakout-playing network from the game's rules that outscored one trained for 500 million steps, 517.4 to 25.3. On Gemma 3 270M, a small model that rarely repeated itself, Astra produced no established math gain. The team released its experiment logs and editing files and announced a benchmark for the method; independent researchers have not yet reproduced the results, and no recursive self-improvement was demonstrated - PauseAI CEO Maxime Fournes warned the work could speed up dangerous progress. The move follows a line of work from ROME (NeurIPS 2022) and MEMIT (ICLR 2023) that allowed targeted edits to facts stored inside a model. In the same week, Anthropic's IPO filing warns that its future models may soon cross its own line for AI-accelerated research, and Artificial Analysis's decision to keep many of its tests secret widened the gap between downloadable models and the leader from 3 points to 12. The signal: the next phase of model improvement is not more training data but one model repairing another, which turns every open-weight deployment into something an AI agent can hot-patch.