AI edits another model's brain directly and lifts math scores sixfold

What's new

Researchers at MIT CSAIL have demonstrated an idea they call "Weight-Space Programming." In short, they let an advanced language model read and write directly in another model's weights. Weights are the numbers inside an AI model that determine how it thinks and responds. Normally, they are changed by training the model for a long time. Here, it happens without gradients and without training.

The researchers used OpenAI's GPT-6 Astra to edit the small open-source model Qwen3.5-0.8B, which is released under the Apache 2.0 license. Astra changed about 0.016 percent of the model's parameters, equivalent to 131,072 weights, according to VentureBeat. The researchers published the results in a blog post on November 2, 2026.

The result: the score on MATH-500, a test of difficult math problems, rose from 7.9 to 48.1 percent, roughly six times higher. On GSM8K, a test of grade-school math problems, it rose from 13.8 to 67.1 percent. On IFEval, which measures whether the model follows instructions, it went from 21.0 to 62.3 percent.

What's clever

Astra didn't just guess. It figured out that the small model's weak math was mainly caused by repetition loops, in which the model repeats text and never arrives at an answer. Astra traced the problem to specific directions in layers 15 and 19 and removed them. The investigation required 50 evaluation requests, 158 conversation turns and 80 Python runs over 20.3 hours, according to VentureBeat.

In other experiments, Astra could tell from raw weights alone which game a small network had been trained on (correct in all 4 out of 4 games), and which language a transformer with 7.16 million parameters had been trained on. It also found all of the attention heads that copy text. In a control task experiment, it fixed a fault by changing 3.5 percent of 171,664 weights, and the success rate went from 0 to 100 percent.

Astra also wrote game-playing networks from scratch based on task descriptions. According to VentureBeat, one of them scored 517.4 in MinAtar Breakout versus 25.3 for a baseline trained with PPO, a widely used training method. In Seaquest, the network reached about 60 percent of PPO's performance, according to 36Kr.

Cheaper or better?

There is no cost analysis against ordinary training, so it can't be said whether the method is cheaper. The price of using GPT-6 Astra is also not stated in the sources for this research, and it is not clear what access costs in Denmark or the EU.

It is better, at least on the selected tests, but with major caveats. According to VentureBeat, the intervention removed an obstacle in the model but did not add new mathematical knowledge. The method also did not work on Gemma 3 270M, because that model did not have one clear failure pattern like repetition loops. It has only been tested on one small model with 0.8 billion parameters. The figures are the researchers' own, the blog post has not been peer reviewed, and the results have not been independently verified. Experiment logs have been posted on GitHub. VentureBeat writes that self-improving cycles have not been demonstrated.

What it's good for

Today, it is first and foremost research. The researchers themselves describe the results as very early and experimental, and it is unclear whether the method works on larger models.

Longer term, the idea points toward a way to debug and fix AI models, almost like fixing a single bug in a program instead of rebuilding it. It may be especially relevant when a small open model fails for a specific and clear reason. For ordinary users and businesses, there is no product to put to use yet, but it is a sign that AI models may in the future be inspected and improved by other AI models.