Changing an AI model usually means training it through repeated mathematical optimization. MIT researchers describe a different approach: GPT-6 Astra examines a model’s internal weights and can modify selected values directly. They call this method “Weight-Space Programming.”
The experiments suggest that a language model can sometimes interpret what a network’s numbers represent and make targeted edits. They also show clear limits: success depended on the task and the model, and not every capability could be created through weight changes.
Reading information in a model’s weights
Weights are numerical values arranged in large tables, or matrices. The researchers gave Astra weights from trained networks and asked it to infer what the networks had learned. In some tests, Astra worked from the numbers alone; in others, it used programming tools and feedback from test runs.
From raw weights, Astra identified which of four computer games a network was meant to play. It recognized basic shapes produced by an image generator, although it mixed up filled shapes and outlines. It also matched the training language of small language models without seeing their training data or running the models.
One test exposed a limit in weight-only inspection. Astra could not initially distinguish which of two fine-tuned language models had been trained to always agree with users. When it could also observe what the models paid attention to while processing text, it identified the sycophantic model across all three pairs.
Writing weights for game-playing networks
Astra also generated weights for networks intended to play three simple computer games. It used the rules and source code, rather than training the networks in the usual way. According to the researchers, its models beat versions trained with a standard method over 100 million rounds in some cases.
Breakout was the strongest reported result: Astra’s model averaged 517 points, compared with 25 for the trained baseline. In two other games, Astra’s models did not match the versions that had undergone the longest training.
The researchers also tried the method on Go played on a 9x9 grid. Astra wrote weights for a network with about 1.5 million parameters. Across 50 test iterations, the network played up to 128 practice games each round, and Astra adjusted its weights based on the results. The resulting model beat a similarly sized model trained on roughly 20 million games.
A small edit improved one language model
The most practical test focused on Qwen3.5-0.8B and its math performance. Astra used a coding environment and 50 test iterations with practice problems, while evaluation used separate problems it had not seen.
Astra found a failure pattern already described in the model documentation: the model could get stuck repeating text instead of reaching an answer. It changed just 0.016 percent of the weights. Afterward, accuracy on the MATH-500 benchmark rose from 7.9 to 48.1 percent, while instruction-following ability increased from 21 to 62 percent.
The change did not eliminate the looping behavior. A third of responses still reached the maximum text length, although that was down from 93 percent before the edit. Astra also could not produce a measurable improvement in Gemma 3 270M, a smaller model that rarely fell into such loops.
Potential, with important limits
The researchers suggest Astra may have removed a blockage in Qwen that had been suppressing capabilities the model already possessed. In this experiment, rewriting weights did not add skills that a model fundamentally lacked. That distinction matters: a targeted edit may help a model use existing abilities, but these results do not show that weight editing can supply any skill on demand.
The approach has drawn both historical comparisons and interest in possible applications. Jürgen Schmidhuber called it “old hat,” pointing to his own work from the 1990s. Khattab responded that those papers were linked and “far ahead of its time,” while arguing that today’s language models are better at translating between human language and a network’s numerical values.
The researchers see possible uses in security audits performed directly on model weights and in changing specific behaviors without expensive retraining. They also acknowledge that the capabilities remain limited, and say a benchmark for measuring them systematically is next on their list. Experiment logs are on GitHub, and the edited Qwen model is on Hugging Face.