MIT CSAIL shows GPT-6 Astra directly reading and editing neural network weights

MIT CSAIL researchers Rikiya Takehi, Jayden Teoh, Alex Zhang and Omar Khattab introduced Weight-Space Programming on 2 November 2026, testing whether frontier language models can inspect, synthesise and repair neural networks directly in parameter space without gradient-based training loops.

Reading raw text dumps of 17,253 MLP weights with zero tools, GPT-6 Astra identified all four Atari-style games (Breakout, Pong, Pac-Man and Freeway) that separate policies were trained to play, and recovered language capabilities and induction heads from stripped transformer middle blocks. Asked to write game-playing networks from scratch with no environment rollouts, Astra produced zero-shot MinAtar weights that outscored policies trained with proximal policy optimisation (PPO) for 100 million steps, and fixed a stuck car-on-ramp PPO policy from 0% to 100% success across 1,024 episodes by editing 3.5% of its weights in one shot.

Given a Python kernel and 50 forward passes (no gradients) on Qwen3.5-0.8B, Astra traced repetitive thinking truncations to MLP neurons in layers 15 and 19 and edited 131,072 parameters (0.016% of weights). Under identical sampling settings, MATH-500 rose from 7.9% to 48.1%, GSM8K rose from 13.8% to 67.1% and out-of-distribution instruction following on IFEval climbed roughly 3 times from 21.0% to 62.3%. On Gemma 3 270M, which rarely looped (8.5%), weight edits could not instil absent mathematical knowledge (6.8% to 7.0%). Edited checkpoints and full agent logs are open-sourced on Hugging Face and GitHub.