'Like gene editing for neural networks': MIT researchers used GPT-6 Astra to edit Qwen's open weights directly, boosting math performance

Open original

MIT CSAIL researchers had an AI model inspect and patch the numerical weights of a small open Qwen model, raising its MATH-500 score from 7.9% to 48.1% by touching about 0.016% of them. The gain came mostly from removing a repetition problem, not from adding new math ability.

  • The edit changed only 131,072 weights in layers 15 and 19, cutting responses that hit the token limit from 93% to 33%.
  • Finding that patch was expensive: the strongest run took 20.3 hours, 80 Python runs and 50 evaluation requests, including a 4.7-hour outage wait.
  • A control on Gemma 3 270M showed no math gain, which suggests the method repairs flaws rather than adding knowledge.
  • Critics split between doubting the reasoning claims and warning that the same tools could speed up capability development.

MIT CSAIL researchers Rikiya Takehi, Jayden Teoh, Alex Zhang and Omar Khattab used OpenAI's GPT-6 Astra to directly edit the weights of Qwen3.5-0.8B, a small open-source model, and their write-up (a research blog post dated November 2) calls the method "weight-space programming." The headline result is that MATH-500 accuracy rose from 7.9% to 48.1%, GSM8K went from 13.8% to 67.1%, and IFEval instruction-following went from 21.0% to 62.3%. The edit touched roughly 0.016% of the weights, specifically 131,072 of them in selected components of layers 15 and 19. The gain came mostly from removing a numerical direction tied to repetition. Before the edit, the model looped on text until it ran out of output tokens (93% of MATH-500 responses hit the limit); afterward that dropped to 33.1%. So the model was not learning more math. It was getting unstuck from a behavior that kept it from using math it already had. The effort behind that patch was not small. The strongest attempt used 50 evaluation requests, 158 conversational turns, 80 Python runs and 20.3 hours of wall-clock time, including 4.7 hours lost to a service outage. The researchers also reported less dramatic but telling experiments. Astra correctly identified four Atari-style games (Breakout, Pong, Pac-Man, Freeway) from weights alone, even though the networks shared architectures, inputs and action sets. In a simulated driving task, a controller that kept getting stuck at a ramp was patched with one failed trajectory and a scene description, changing 3.5% of its weights for a reported 100% success rate. In the MinAtar Breakout test, a network Astra built from the game's rules and source code scored 517.4, against 25.3 for a PPO-trained network given 500 million steps. Astra trailed the longer-trained PPO baselines in Seaquest and was slightly behind in Asterix, though it beat the 100-million-step runs on all three games. Reactions split as you'd expect. Databricks' Yuchen Jin and Voyage AI's Chandan Singh praised the idea as a possible route to cheaper maintenance and improvement without conventional retraining, and Singh floated models editing other models or themselves. The authors themselves say they have not shown savings in compute, speed or cost, and that direct editing did not clearly beat conventional training. Pause AI CEO Maxime Fournes argued the approach could speed up risky development, and a skeptical commenter questioned whether any of this reflects better underlying reasoning. Both points are worth holding onto: the result is a targeted repair demonstrated on one small model, with generalization and savings still unproven.

Ask about this article