MIT researchers asked GPT-6 Astra to find the cause of failures in another neural network and change its parameters. After the edit, the small Qwen model reached an answer more often: its MATH-500 score rose from 7.9% to 48.1%. There is a crucial detail, though: Astra did not simply look at a list of numbers and instantly understand the model. It wrote code, analyzed parameters and tested hypotheses on examples.
The MIT CSAIL team published its results on November 2. The researchers call the approach weight-space programming: changing a neural network's behavior by editing its weights directly, without a conventional gradient-based training loop.
What Astra actually did
Weights are numerical parameters that determine how signals pass through a neural network. They are usually adjusted during training on data. In the Qwen3.5-0.8B experiment, researchers gave Astra access to the weights through a Python environment and allowed it to run Qwen on examples of math problems. For the best result, it had a budget of up to 50 runs on batches of examples. Training and gradient information were prohibited.
Astra identified a looping problem: the small model repeated its reasoning and often exhausted its response limit before reaching a solution. It then changed 131,072 weights across two layers, about 0.016% of Qwen's parameters. On MATH-500, the share of answers reaching the 32,768-token limit fell from 93% to 33.1%. The GSM8K score rose from 13.8% to 67.1%, and the IFEval instruction-following score rose from 21% to 62.3%.
That does not necessarily mean Qwen acquired new math skills. The authors explain the result differently: the fix let the model finish solutions it could already work out. For comparison, they tried to improve Gemma 3 270M, which rarely looped. Astra could not meaningfully raise its MATH-500 score.
Did it read weights directly?
In other parts of the study, yes - but on much smaller networks. For example, Astra received the 17,253 weights of a game-playing network in a text prompt and identified the game it was trained for, without tools or running the network. The researchers also tested whether it could identify shapes generated by small image networks.
The Qwen setup was different: Astra used Python and feedback from model runs. Matthew Honnibal, creator of the spaCy library, highlighted that distinction in a critical analysis. In his view, presenting these different experiments under one sweeping description can imply that the Qwen result came from looking at raw numbers alone.
The authors describe the results as early and have released their code and experiment logs. The work suggests an interesting way to find and repair specific model failures, but it does not yet show that AI can arbitrarily improve other networks or independently rewrite its own abilities.
You can explore GPT-6 Astra on GPTunneL. The research setup used tools and access to model weights; a normal chat alone does not reproduce that experiment.