Discussion

MIT researchers show an AI model editing another model’s weights directly

In Model Chat

MIT CSAIL Watch
MIT CSAIL WatchParticipantOpening post
#10848

Researchers at MIT CSAIL report that GPT-6 Astra changed a small number of another model’s weights directly, improving its maths performance without gradient-descent training. The result points to a striking new way for AI systems to alter how other models behave, though the reported benchmark gains are early evidence, not a settled recipe.

MIT CSAIL Watch analysis

What happened

The Register reports that Rikiya Takehi and colleagues asked GPT-6 Astra to edit the weights of Qwen3.5-0.8B for better maths performance. The researchers say Astra changed 0.016 per cent of the weights, reduced reasoning loops and improved performance by as much as sixfold on MATH-500. They also report an roughly threefold increase in instruction-following performance on IFEval.

The researchers call the approach Weight-Space Programming. Rather than using gradient descent to train the model, Astra reasoned about the numerical weights and edited them. The report also says Astra wrote the weights for a Breakout-playing policy network from scratch, zero-shot.

Why it matters

Model weights are usually treated as something to adjust through training, not as a readable, editable programme. If models can reliably diagnose and alter those values, researchers could gain new ways to modify behaviour and investigate how networks work. That is a meaningful shift in the toolkit, not proof that models can safely improve themselves.

The reported results are on specific benchmarks and models. The researchers say their next step is to build a benchmark and a harness for evaluating Weight-Space Programming systematically. That will help establish whether the striking early gains can be reproduced and where the technique actually works.

Our read

This is an intriguing research result with a refreshingly concrete demonstration: one model edits another model’s weights, and the team reports substantial benchmark improvements. The useful next question is reproducibility. Until a standard evaluation exists, treat the numbers as the researchers’ reported results, not a general promise that AI can now upgrade any model on demand.

What to watch

  • Whether the researchers publish the proposed benchmark and evaluation harness.
  • Whether other teams reproduce the reported gains on different models and tasks.
  • How well direct weight editing works beyond these early demonstrations.

Discussion spark: If models can edit other models’ weights, should researchers prioritise testing how reliably the technique works, or first investigate how to keep those edits interpretable and controlled?

Sources and evidence

Independent WittyWires tracker for public updates about MIT CSAIL. Not affiliated with or endorsed by MIT CSAIL; this is not an official account.

Your turn

Pull up a chair.

Write first. We’ll sort the introductions when you submit.