MIT Researchers Say GPT-6 Astra Can Edit AI Models by Changing Their Weights

Experiments described by MIT researchers suggest GPT-6 Astra can inspect and alter other models’ weights without conventional retraining. In one test, changing 0.016% of a Qwen model’s weights raised its score on a math benchmark, though the method did not work across all models or tasks.

MIT researchers report that OpenAI’s GPT-6 Astra can analyze and modify the numerical weights inside other AI models, a process they call “Weight-Space Programming.” Rather than retraining a model through conventional optimization, Astra works directly with its weights, sometimes using coding tools and feedback from test runs.

In experiments, Astra inferred what several trained networks had learned by examining their weights. It identified which of four games a network was built to play, recognized the basic shapes produced by an image generator, and matched the training language of small language models. It could not distinguish two language models fine-tuned to agree with users from weights alone, but succeeded when it could also observe which parts of text the models attended to.

The researchers also asked Astra to create weights for models that played simple games, using game rules and source code rather than training data. In Breakout, its model averaged 517 points, compared with 25 for a conventionally trained model. Astra’s models did not outperform the longest-trained versions in two other games. For 9x9 Go, a network Astra adjusted through 50 rounds of testing beat a similarly sized model trained on about 20 million games.

The most notable results came from an attempt to improve the math performance of Qwen3.5-0.8B. Using a coding environment and 50 rounds of testing, Astra changed 0.016% of the model’s weights. Its accuracy on the MATH-500 benchmark rose from 7.9% to 48.1%, while its instruction-following score increased from 21% to 62%.

The edit also reduced, but did not eliminate, a behavior in which Qwen repeatedly generates text without reaching an answer: about a third of responses still reached the maximum text length, compared with 93% beforehand. Astra produced no measurable improvement when tested on the smaller Gemma 3 270M model, which rarely showed the same looping behavior. The researchers suggest the Qwen edit may have removed an obstacle suppressing existing abilities, while noting that rewriting weights did not demonstrate a way to add capabilities a model fundamentally lacks.

The researchers see possible uses in auditing model weights for security issues and making targeted behavior changes without costly retraining. They acknowledge that the approach remains limited and say developing a benchmark for measuring it systematically is a next step. Experiment logs are available on GitHub, and the edited Qwen model is on Hugging Face.

Sources: The Decoder.

Related articles