MIT researchers say GPT-6 Astra can edit other models' weights

The researchers tested the method on reading and writing tasks. In one set, Astra received raw weights from trained networks and was asked what they had learned. Depending on the experiment, it worked directly with lists of numbers or used programming tools and feedback from test runs, the researchers said.

Reading weights

Astra correctly identified which of four computer games a network was built to play, the researchers said. It also matched the training language of small language models without access to training data or the ability to run the models.

Astra also recognized the basic shapes an image generator produced, though it confused filled shapes with outlines, according to the research blog. In another test, it had to identify which of two fine-tuned language models was trained to always agree with users. From weights alone it could not tell at first; only when it could observe the models' attention patterns did it identify the sycophantic model across all three pairs.

Writing weights

The researchers then asked Astra to generate weights for models meant to play three simple computer games, based only on game rules and source code. In Breakout, Astra's model scored an average of 517 points against 25 for a baseline trained over 100 million rounds, they said. In two other games, Astra fell behind the longest-trained versions.

For Go on a 9x9 grid, Astra wrote the weights of a network with about 1.5 million parameters. Given 50 test iterations, with the network playing up to 128 practice games each round and Astra correcting its weights, it beat a comparably sized model trained on roughly 20 million games.

The most practical experiment involved Qwen3.5-0.8B. Astra's task was to improve the model's math skills using a coding environment and 50 test iterations with practice problems. Astra identified a thinking-loop problem that Qwen's own documentation warns about, then changed just 0.016 percent of all weights.

On the MATH-500 benchmark, accuracy jumped from 7.9 percent to 48.1 percent, the researchers said. Instruction-following ability improved from 21 percent to 62 percent. The thinking loops were not fully eliminated: a third of responses still hit the maximum text length, down from 93 percent before.

What it means

The researchers see potential for security audits directly on model weights and for targeted behavior changes without expensive retraining. They also acknowledge limits. With Gemma 3 270M, a smaller model that rarely fell into such loops, Astra could not produce any measurable improvement, they said. They suspect Astra removed a blockage in Qwen that was suppressing existing capabilities; skills a model fundamentally lacks could not be added by rewriting individual weights in this experiment.

Separately, on Artificial Analysis' Intelligence Index V5, GPT-6 Astra (max) scores 57, two points behind Google's Gemini 4 Carbon (high) at 59, Tom's Hardware reported. That general ranking measures broad model capability, not the narrower ability to inspect and edit other models' weights that the MIT experiments test.

Next steps

A benchmark for systematic measurement is next on the researchers' list, they said. The experiment logs are on GitHub, and the edited Qwen model is on Hugging Face.