GPT-6 Astra 通过权重直接改写其他 AI 模型

The usual training process of step-by-step mathematical optimization isn't needed. Depending on the experiment, Astra worked directly with lists of numbers or used programming tools and feedback from test runs.

The researchers call the approach "Weight-Space Programming": instead of expensively retraining an AI model, a language model acts like a mechanic turning individual knobs.

Astra identifies games just by looking at a model's raw numbers

In several experiments, the researchers fed Astra the raw weights of trained networks and asked what those networks had learned. Astra correctly identified which of four computer games a network was supposed to play. It also recognized the basic shapes an image generator produced, though it confused filled shapes with outlines. It correctly matched the training language of small language models too - without access to training data and without the ability to run the models.

In another test, Astra had to figure out which of two fine-tuned language models was trained to always agree with users. From the weights alone, it couldn't tell the difference at first. Only when Astra was also allowed to observe what the models paid attention to during text processing did it correctly identify the sycophantic model across all three model pairs.

Models written from scratch beat trained competitors

Astra went even further: It generated weights for models meant to play three simple computer games, entirely without training, based only on the game rules and source code. According to the researchers, these models outperformed ones trained with a standard method over 100 million rounds. In Breakout, Astra's model scored an average of 517 points versus 25 for the trained baseline. In two other games, Astra fell behind the longest-trained versions.

For the board game Go on a 9x9 grid, Astra wrote the weights of a network with about 1.5 million parameters. Given a budget of 50 test iterations where the network played up to 128 practice games each round and Astra then corrected its weights, it beat a comparably sized model trained on roughly 20 million games.

Targeted fixes instead of full retraining

The most practical experiment involved the language model Qwen3.5-0.8B. Astra's task was to improve its math skills, using a coding environment and 50 test iterations with practice problems. Evaluation happened on separate problems Astra had never seen.

Astra identified a problem that Qwen's own model documentation already warns about: thinking loops where the model endlessly repeats the same text instead of reaching an answer. Astra then changed just 0.016 percent of all weights. The result: accuracy on the MATH-500 benchmark jumped from 7.9 to 48.1 percent. Instruction-following ability improved from 21 to 62 percent. The thinking loops weren't fully eliminated, though, as a third of responses still hit the maximum text length, down from 93 percent before.

With a smaller model (Gemma 3 270M) that rarely fell into such loops, Astra couldn't produce any measurable improvement.

The researchers suspect Astra was able to remove a blockage in the Qwen model that was suppressing existing capabilities, but skills that a model fundamentally lacks couldn't be added by just rewriting individual weights in this experiment.

Old news or new paradigm?

AI pioneer Jürgen Schmidhuber commented on X that the approach was "old hat," pointing to his own work from the 1990s. Khattab responded that they link to those papers, which were "far ahead of its time." The new aspect, he said, is how well today's language models can translate between human language and a network's numerical values.

The researchers see potential in their approach for things like security audits directly on model weights or targeted behavior changes without expensive retraining. They also acknowledge that the capabilities are still limited. A benchmark for systematic measurement is next on their list. The experiment logs are on GitHub, and the edited Qwen model is on Hugging Face.