MIT researchers show AI models can edit other models' weights
MIT CSAIL researchers demonstrated that frontier AI models can directly modify the weights of other neural networks to alter their behavior.
MIT CSAIL researchers found that frontier AI models, such as OpenAI's GPT-6 Astra, can rewrite the weights of another model to change its behavior. In one experiment, Astra edited the weights of Qwen3.5-0.8B to improve its math capabilities. The model changed only 0.016 percent of the weights, raising the MATH-500 benchmark score from 7.9% to 48.1%. Astra also wrote a neural network from scratch to play Breakout. The authors expect this Weight-Space Programming approach to enable recursive self-improvement and reduce model sycophancy.
Critics on Hacker News argued the findings are misleading. One user noted that the Qwen experiment used Python programs rather than raw weight inspection. spaCy author Matthew Honnibal stated the paper conflates capabilities requiring code execution with those using raw weights. Sean Goedecke said the update process mirrors standard human ML research steps. Another user claimed a local Qwen model can interpret weights similarly.
The next step is building a benchmark and harness for Weight-Space Programming. The researchers believe this will turn the process into a rigorous, reproducible engineering discipline. The paper was published on 3rd November 2026, with code available in the RikiyaT/lm_understand-logs repository.
Sources
- The RegisterForget retraining - AI can now tweak another AI's weights ↗
- Hacker News: front pageThe 'Astra can read and write weights to improve Qwen' post is misleading ↗
- Hacker News: front pageQwen 3.8 27B on my laptop can interpret model weights, too ↗
Every sentence of this story was checked against facts quoted from these articles. How we make the news
Spotted an error? Tell us
