GPT-6 Astra Directly Adjusts Weights, Boosting 0.8B Model's Math Performance by 6x
Some researchers claim that by having GPT-6 Astra directly edit the model weights of Qwen3.5-0.8B—altering just 0.016% of the parameters—they fixed the smaller model's reasoning loops. This resulted in up to a 6x improvement in math capabilities and a 3x boost in instruction following, all without gradients or training.
GPT-6 Astra Directly Modifies Weights, Boosting a 0.8B Model's Math Performance by 6x
Some researchers claim that by having GPT-6 Astra directly edit the model weights of Qwen3.5-0.8B—altering just 0.016% of the parameters—they were able to fix the smaller model's reasoning loops. This resulted in a performance boost of up to 6x in mathematics and 3x in instruction following, all without requiring gradients or any training.
The authors call this method "weight space programming," suggesting that state-of-the-art large models can now natively read and write neural network weights. The experimental details have been published on their blog.
Rikiya Takehi(@rikiyatakehi): https://x.com/rikiyatakehi/status/2117313742891327994
