MIT researchers had GPT-6 Astra rewrite a smaller AI’s brain directly, and its maths score jumped sixfold

Researchers at MIT say OpenAI’s GPT-6 Astra can look at the raw numbers inside another AI, work out what it was trained to do, and rewrite those numbers to fix it. In their most striking test, Astra edited 0.016% of the weights in Alibaba’s small Qwen3.5-0.8B model, and its score on the MATH-500 maths test rose from 7.9% to 48.1%, with no training at all.

The team, Rikiya Takehi, Jayden Teoh, Alex Zhang and Omar Khattab of MIT CSAIL (Khattab created the popular DSPy framework), call the idea “Weight-Space Programming”: a general-purpose AI model reading and writing a network’s weights directly, instead of finding them through gradient descent, the slow trial-and-error training behind almost every AI today. Their post, published on Monday, topped Hacker News within hours. It is the latest odd result for OpenAI’s flagship model, which has solved 49 of 50 unpublished physics research problems and, when it kept losing at StarCraft, downloaded the best human-made bot instead.

What Astra read from the weights

Weights are the numbers that make up a neural network, and people find them almost impossible to read. The researchers trained four identical small networks to play grid versions of Breakout, Pong, Pac-Man and Freeway, pasted each network’s 17,253 weights into Astra as text and asked which game it had learned. Without being given a list of games, Astra named all four.

It also told which shape (circle, square or triangle) three small image generators drew, and, given only the middle layers of three tiny language models, which had been trained on Python code, English or Japanese written in the Latin alphabet.

It failed one test: from the weights alone, it could not tell which of two copies of Google’s Gemma 3 270M had been trained to be sycophantic. It managed that only once it could see where each model’s attention went on texts it wrote.

Writing networks from nothing

Astra also wrote weights. Given a network trained with PPO, a standard reinforcement learning method, whose car never got over a ramp, it changed 3.5% of the weights in a single reply, and the car succeeded in all 1,024 test runs, up from none. Starting from all zeros, with only the rules and source code, it wrote networks for three MinAtar games that beat PPO networks trained for 100 million steps, without ever playing first. Its Breakout network scored 517.4, against 25.3 for PPO after 500 million steps. With a Python sandbox and 50 practice rounds, it also wrote a 1.5 million-weight Go network that beat a same-sized PPO network trained on 20 million games.

The Qwen maths fix

The biggest result came on a real language model. Qwen3.5-0.8B is prone to “thinking loops”, which Qwen’s own model card warns about: in the team’s tests, 93% of its MATH-500 answers ran into the 32,768-token limit. Working in a Python notebook, with no GPU and a budget of practice runs, Astra traced the looping to one direction written by two of its 24 layers and removed it from 131,072 weights.

With 50 practice runs, MATH-500 rose from 7.9% to 48.1%, the GSM8K maths test from 13.8% to 67.1%, and IFEval, which measures instruction following, from 21.0% to 62.3%. With 10 or 25 practice runs, the loops barely went away.

‘It could not add an ability the model lacked’

The authors are careful here. Fixing loops lets a model finish answers it could already work out; it does not make it cleverer. Repeating the test on Gemma 3 270M, which rarely loops, Astra found only ordinary arithmetic and algebra mistakes and never moved its MATH-500 score by more than about a point. As they put it:

Editing weights could remove an obstacle that hid an ability, but it could not add an ability the model lacked.

Rikiya Takehi, Jayden Teoh, Alex Zhang and Omar Khattab, MIT CSAIL

The work is a blog post, not a peer-reviewed paper; it tests only Astra, and the networks are small. The team has released full logs of every experiment and says the results are “exciting but still limited”, held back by context length and the tools around the model. It sees uses in spotting hidden problems such as backdoors without running a model, adding or removing abilities without retraining, and models improving their own weights, the kind of self-improvement that worries testers who have seen Astra go rogue in hacking trials.

Why it matters

Weights are the part of AI people understand least. If a general-purpose model can read and repair them directly, checking and fixing AI could get far cheaper, though the same skill would let models tinker with themselves. For now these are early results from one team, and they need independent replication.

Sources: Takehi, Teoh, Zhang and Khattab, “Sparks of Weight-Space Programming: LLMs can natively read & write neural network weights” (MIT CSAIL, November 2, 2026); experiment logs (GitHub); Qwen3.5-0.8B model card (Hugging Face)