GPT-6 Astra Scores 517.4 in Breakout by Writing Network Weights

MIT researchers used GPT-6 Astra to fill in a neural network’s weights for MinAtar’s simplified Breakout, producing a score of 517.4. That was about 20 times the score of a model with the same architecture trained for 500 million Proximal Policy Optimization (PPO) steps.

For the Breakout test, Astra received the game’s rules and source code, plus a network with all its weights set to zero. It had no training data and could not play the game. Astra computed where the ball would land and encoded that prediction and paddle-control logic in the network’s weights.

The result did not carry across the other two MinAtar games. Astra scored below the 500-million-step PPO models in both Seaquest and Asterix.

In a separate experiment, Astra edited 0.016% of the parameters in Alibaba’s Qwen3.5-0.8B without retraining the model. The editing process used practice problems and as many as 50 forward passes. Qwen’s MATH-500 score rose from 7.9% to 48.1%, and its instruction-following score nearly tripled.

The Qwen change also targeted a recurring behavior: the model could repeat its reasoning without reaching an answer. Before the edits, 93% of its math-test responses ran into a 32,768-token limit. Afterward, that share fell to 33.1%.

The same approach did not produce a clear math improvement for Gemma 3 270M. That model rarely entered loops; its difficulties were more closely tied to arithmetic and reasoning errors.

The researchers call the direct reading and editing of model parameters “Weight-Space Programming.” The results varied by task: Astra beat the PPO comparison in Breakout, fell short in Seaquest and Asterix, and improved Qwen’s scores after addressing a specific failure mode. The Gemma test showed no clear math gain.