MIT researchers demonstrated a method called Weight-Space Programming, where OpenAI's GPT-6 Astra modifies the internal weights of other AI models without traditional retraining. The system analyzes raw numerical values to identify model functions, such as determining which computer game a network was trained to play.
In tests, Astra generated weights for models playing simple games based on rules and source code. These models outperformed standard training methods in some cases, such as in Breakout, where Astra's model scored an average of 517 points compared to 25 points for the baseline.
For the board game Go on a 9x9 grid, Astra created a network with 1.5 million parameters. With a budget of 50 test iterations, it outperformed a model trained on 20 million games. The approach also improved the math skills of the Qwen3.5-0.8B model by modifying 0.016 percent of its weights.
Following these adjustments, the accuracy of the Qwen model on the MATH-500 benchmark increased from 7.9 to 48.1 percent, and instruction-following improved from 21 to 62 percent. Researchers suggest the method could be used for security audits or targeted behavior changes in AI systems.
