A small network of charcoal spheres and rods with one copper repair and a precision tool beside it.
A targeted repair is one way to investigate what a model is doing.Sint

Breakdown / AI research · Model evaluation

AI can edit another model's weights. Here's what that achieved.

Researchers used GPT-6 Astra to repair a smaller model's repetitive thinking. A failed follow-up helps explain what the result does, and doesn't, show.

A small AI model keeps repeating itself while trying to solve a maths problem. It runs out of room before producing an answer. Could another AI find the cause and change the model itself?

MIT CSAIL researchers Rikiya Takehi, Jayden Teoh, Alex Zhang and Omar Khattab report doing that with GPT-6 Astra and the smaller Qwen3.5-0.8B model. Astra edited a tiny fraction of Qwen's weights, the learned numerical settings inside a neural network. Its score on the MATH-500 maths test rose from 7.9% to 48.1%. Research report

The useful detail is what Astra repaired. A model that gets stuck repeating itself can miss a problem it otherwise has the ability to solve.

Reading weights and repairing a model are different jobs

The researchers gave Astra Python tools, practice problems and a limited budget for running Qwen. It compared looping and non-looping behavior, identified weights contributing to repetition and reduced their influence. Feedback from running the model helped guide the repair. Experiment setup

A follow-up on Gemma 3 270M, which rarely looped, produced no established improvement. Its arithmetic and reasoning mistakes had no single fault the edits could remove. We haven't reproduced either experiment. Follow-up and limitations

Think of fixing a calculator's stuck key: removing the obstruction lets it finish a calculation. Teaching it a new operation is a different problem. The analogy isn't a description of neural networks, but it captures why these two results belong together.

A laptop experiment shows another part of the idea

The work also includes smaller tests of whether an AI can infer a neural network's behavior from its weights without running it. Developer Simon Willison tried the authors' game-identification prompts with Qwen 3.8 27B on his own laptop.

His model identified Breakout, Pong and Pac-Man, but mistook Freeway for a Pac-Man-like game. The successful attempts took roughly 46 minutes to an hour and five minutes each. This was a separate developer's attempt on the researchers' examples, rather than a test on new, unseen networks. Willison's account

Software engineer Sean Goedecke offers a useful comparison. He shows how plotting the strength of a game network's input connections can reveal where it pays attention on the board. He argues that the results can be understood through techniques a human researcher could use. That is a methodological critique, not a measured human-versus-AI contest. His analysis

The interesting prospect is having an assistant help investigate and repair a model in ways a person can inspect. To judge that prospect, we'd want new examples, all the failed attempts, the time and computing used, and checks beyond the behavior being repaired. The next convincing repair will need to work beyond the example that suggested it.

Sources