MIT researchers say GPT-6 Astra can read and rewrite neural network weights

On 2 November 2026 four MIT CSAIL researchers, Rikiya Takehi, Jayden Teoh, Alex Zhang and Omar Khattab, reported that GPT-6 Astra, OpenAI’s top-tier model, can work directly on another network’s weights, the numbers a neural network learns in training. In what they call weight-space programming, the model reads those numbers, reasons about what they compute and writes new ones, instead of leaving the job to gradient descent, the usual training method.

What the experiments showed

In most tests the networks were small and Astra could not run them.

• Reading: from the raw weights alone, with no list of options, Astra named the game each of four small networks was trained to play. It also identified what three image generators drew and, from the middle layers alone, which language each of three small language models had learned.
• Fixing: a network trained to drive a car over a ramp in a physics simulator always stalled. In one reply, Astra worked out that the car needed a run-up and changed 3.5% of the weights; the car then succeeded in all 1,024 test runs.
• Writing: for three simple MinAtar games, Astra wrote networks from scratch that beat networks trained for 100 million steps, though not always those trained for 500 million.

Fixing a small model’s loops

The largest test used Qwen3.5-0.8B, an open model whose model card warns it is prone to ‘thinking loops’. In a Python session, with no training and 50 trial runs on practice problems, Astra traced the looping to one direction in two layers and subtracted it, changing 0.016% of the weights. Answers that ran to the length limit fell from 93% to 33%, and the authors report that MATH-500 rose from 7.9% to 48.1% and IFEval, an instruction-following test, from 21.0% to 62.3%.

Astra could not improve Gemma 3 270M, a smaller model that rarely loops. Nor could it tell from the weights alone which of two Gemma models had been trained to agree with the user, though it could once shown their attention patterns.

Caveats

This is a blog post, not peer-reviewed research, and every result is the authors’ own. They have published full logs and the edited model so others can check. Only GPT-6 Astra was tested, on small networks. A separate preprint posted on 5 November found that Claude Opus 5.5’s readings of small image classifiers leaned on familiar public datasets, calling them ‘local arithmetic plus dataset priors’ rather than tracing how a network computes.

There is nothing to buy yet. GPT-6 Astra costs US$10 per million input tokens and US$50 per million output tokens in OpenAI’s API, and the authors say they are building a benchmark for the technique.