SignalTech/AI

Astra can read and write weights to improve Qwen

Astra modifies Qwen by reading and writing model weights.

Context

What happened: Researchers report that frontier LLMs can natively read and write neural network weights. GPT-6 Astra identified which game four PPO-trained networks played, named what three DDPMs generate, and edited Qwen3.5-0.8B weights to fix reasoning loops. Key numbers: Astra edited 0.016% of Qwen3.5-0.8B weights, raising MATH-500 from 7.9% to 48.1% and tripling IFEval. Why it matters: The authors call this Weight-Space Programming and argue that LLMs reasoning about weights directly could replace symbolic optimizers like Adam or Muon for optimization and interpretability.

What changed

Researchers report that frontier LLMs can natively read and write neural network weights. GPT-6 Astra identified which game four PPO-trained networks played, named what three DDPMs generate, and edited Qwen3.5-0.8B weights to fix reasoning loops.

Why it matters

The authors call this Weight-Space Programming and argue that LLMs reasoning about weights directly could replace symbolic optimizers like Adam or Muon for optimization and interpretability.

Sources