无梯度优化
Methods that adjust model parameters to improve performance through agent reasoning or search strategies, without relying on bac
Simon Willison verified that Qwen 3.8 27B running locally on a laptop can perform 'Weight-Space Programming,' interpreting and modifying raw neural network weights directly from text. This confirms the capability is not exclusive to frontier models like GPT-6 Astra, demonstrating that mid-sized open-source models can also reconstruct or optimize networks without gradient feedback.
Frontier LLM Astra demonstrates native capability to read and write neural network weights, optimizing small models without gradients. By editing just 0.016% of Qwen3.5-0.8B's weights, Astra boosted its MATH-500 score from 7.9% to 48.1% and tripled IFEval performance. This 'Weight-Space Programming' paradigm may reshape model optimization and interpretability research.