Skip to content- Sources: primary, discussion
- Summary: The authors report that GPT-6 Astra identified which game small PPO policies were trained on from raw weights, found induction heads in a character transformer without running it, and wrote MinAtar policies in one shot that beat PPO trained for 100 million steps. Editing 0.016% of Qwen3.5-0.8B weights to fix reasoning loops raised MATH-500 from 7.9% to 48.1% and tripled IFEval. They call the approach Weight-Space Programming.
- Why it matters: If it holds up, LLM-driven weight edits become a tool for interpretability and targeted model fixes alongside fine-tuning.
- Comments: Results are early and from a blog post, not a peer-reviewed paper, and the larger experiments used a Python sandbox and limited forward passes.