Patient: Qwen3.5-0.8B · Surgeon: GPT-6 Astra
An AI edited another AI's weights, and the math score jumped sixfold. Here's what actually happened.
Researchers at MIT CSAIL asked one AI model to repair another by changing its numbers directly, with no training run. It worked on one model and barely moved a second, and the difference is the whole story.
The short version
GPT-6 Astra changed 0.016% of the weights in a small model called Qwen3.5-0.8B. Its score on a math test rose from 7.9% to 48.1%. Qwen had been stuck in reasoning loops, and the edit removed the cause, so it didn't learn new math. It's a real case of AI debugging AI, and it still needs independent replication.
What happened
On November 2, 2026, Rikiya Takehi, Jayden Teoh, Alex Zhang and Omar Khattab of MIT CSAIL published a post on what they call weight-space programming. Weights are the numbers that make up a neural network. Normally they're set by training, a long automatic process. Here, a language model reads and changes them on purpose.
The post runs eight experiments. The one that spread is the last: asked to make Qwen3.5-0.8B better at math, Astra worked in a Python environment with the weights, no GPU and a small budget of test runs on practice problems.
How the edit worked
Qwen's model card already warned about looping. On the MATH-500 test, 93.0% of its answers ran to the 32,768-token limit, so most never reached an answer. Astra traced the looping to one direction written by two internal layers, 15 and 19, and subtracted that direction from 64 output columns in each. That's 131,072 weights.
Give Astra a bigger budget
A forward pass is one run of the model on an input. Pick how many Astra was allowed.
Scores from the authors' table. Math and instruction scores are higher when better. The token limit share is lower when better. The big jump only comes at 50.
What it means, and what it doesn't
What it means: a frontier model did a piece of interpretability research, the study of what's going on inside a network, and acted on it. People online are calling this a step toward recursive self-improvement, meaning AI doing the work of building better AI. That framing isn't the authors' headline claim: they list self-improvement as a possible future use and call the results "very early".
What it doesn't mean: Astra didn't improve itself, and Qwen didn't gain a skill. The authors tried the same thing on Gemma 3 270M, which rarely loops. Its MATH-500 score went from 6.8% to between 7.0% and 7.7%. In their words, editing "could not add an ability the model lacked". It isn't takeoff.
The strongest objection: the word "natively". For this experiment Astra used code and a known kind of technique, and the post itself compares the fix to deleting a single direction from a model. So the Qwen result doesn't show an AI reading weights by eye. Other experiments in the post are stranger: with no tools, Astra named the game behind each of four small networks from 17,253 raw weights, and it wrote a Breakout network from scratch that scored 517.4, against 25.3 for one trained for 500 million steps. It also has clear failures: from weights alone it failed all four attempts to spot a sycophantic model.
Reactions
Astra used code to perform some standard interp techniques, found a looping direction, and ablated it from the weights.Adam Karvonen, @adam_karvonen, November 2
Waiting for replication but this is a pretty big deal.Ethan Mollick, Wharton professor, @emollick, November 2
Questions people are asking
Did the math test leak into the practice problems?
The post says Astra practiced on problems from GSM8K and MATH that aren't in the test sets, and it was scored on benchmarks it never saw.
Could an edit like this break other things?
Instruction following improved too, from 21.0% to 62.3%, but the post reports benchmarks rather than a full regression test of everything the model does.
Is this peer reviewed?
It's a blog post, and the post doesn't mention a paper. The prompts, outputs, code and evaluations are in a public repository, so others can check them.
What comes next?
The authors say they're building a benchmark and a harness for this kind of work. They name context length and harness inefficiency as today's limits.
Updates
November 2, 2026. This page went up. It covers everything known as of November 2, 2026.