Weight-Space Programming explores reading and writing model weights directly
What happened
On November 3, Latent Space reported on Weight-Space Programming research, which explores having frontier large language models inspect, generate and modify neural network weights directly from natural-language goals, without a gradient-descent training loop. Demos described in the report include identifying small-model behavior, generating game strategies, and editing the weights of Qwen3.5-0.8B. In that case, modifying 0.016% of the model's weights raised its MATH-500 score from 7.9% to 48.1%. The progress the report presents is these research demos and their test results.
Written by AI from the coverage · updated 2 hours ago
Coverage
Follow the reports to see the story from different sides.
- Latent Space[AINews] LLMs can read & write neural network weights: "Weight Space Programming" cuts Autoresearch out of the RSI loop
Weight-Space Programming 研究探索让前沿 LLM 根据自然语言目标直接检查、生成和修改神经网络权重，无需梯度下降训练循环。报告的演示包括识别小模型行为、生成游戏策略，以及编辑 Qwen3.5-0.8B 的 0.016% 权重，将 MATH-500 成绩从 7.9% 提升至 48.1%。
Heat over time
Not enough continuous observations to draw a trend yet.