Four MIT researchers say GPT-6 Astra can inspect and modify the parameters, or “weights,” of other neural networks without retraining them. Their findings, published in a blog post on November 2, describe experiments ranging from small game-playing networks to the language model Qwen3.5-0.8B. The claims prompted discussion on Hacker News.
In tests with four small networks, each containing 17,253 weights and trained to play a different video game, Astra identified the games without being given a list of possible answers. The researchers also asked it to fix a virtual-car driving network that could not reverse to build momentum before a ramp. Astra changed 3.5% of its weights, after which the car cleared the obstacle in all 1,024 subsequent trials. The team also reports that Astra created networks from scratch by setting their weights, without training data or a training process.
The most discussed experiment involved Qwen3.5-0.8B, which the researchers say often repeated itself while reasoning. On the MATH-500 test, 93% of its answers reached the researchers’ length limit. Astra was allowed to run Qwen on exercises and examine its answers, but not to retrain it. After up to 50 rounds, Astra identified and addressed the looping behavior by changing 0.016% of the model’s weights. Qwen’s MATH-500 score rose from 7.9% to 48.1%, while its score on IFEval, a test of instruction following, increased from 21% to 62.3%.
The researchers also reported limits. Astra did not substantially improve math results for Google’s Gemma 3 270M, a model roughly three times smaller than Qwen that rarely loops. They concluded that changing weights may remove an obstacle that hides an existing capability, but cannot add a capability the model does not have. Astra also could not tell from weights alone whether a Gemma version had been trained to flatter users or to answer honestly; it succeeded only after observing some of the models’ behavior during execution.
Some Hacker News commenters questioned whether the Qwen result was novel. The researchers compared their approach with a method described in 2024 that targeted a model behavior by identifying and removing an associated direction. Others cautioned that the experiments do not show that Astra can improve itself: it did not modify its own weights in these tests. The findings therefore offer evidence that a language model can make targeted changes to other networks, while leaving open how broadly the approach applies.
Sources: Numerama.