Reproduce “Sparks of Weight-Space Programming: LLMs can natively read & write neural network weights” on OpenResearch

About this paper

Rikiya Takehi, Jayden Teoh, Alex Zhang, Omar Khattab

The authors show that GPT-6 Astra, a general-purpose frontier LLM, can read and write neural network weights directly, without gradient descent or trained helper models. Across eight experiments it identifies what small networks were trained on from their weights, fixes and writes game-playing policies, and edits 0.016% of Qwen3.5-0.8B's weights to stop reasoning loops. They call this Weight-Space Programming.

Key insights

• Astra reads structure directly: it named the game for all four PPO policies and found every induction head in transformers stripped of their input and output layers.
• It writes networks as programs, e.g. a Breakout layer that computes where the ball will land and a Go network built from liberty and group-danger features.
• Weight edits can remove an obstacle that hides an ability, like Qwen's looping, but could not add math ability to Gemma 3 270M, which rarely loops.
• From weights alone it could not identify a sycophantic Gemma 3 270M; with attention patterns on its own probe texts it picked the sycophant in all three pairs.

Summary by alphaXiv.