Primary sourceelvis
View sourceEvent ·Multi-source
研究称 LLM 可直接改写神经网络权重,GPT-6 Astra 微调 Qwen3.5-0.8B 数学能力提升至 6 倍
研究团队让 GPT-6 Astra 直接编辑 Qwen3.5-0.8B 的权重来改进数学能力,全程不使用梯度或训练。模型只修改了 0.016% 的权重,就把 0.8B 模型的数学能力提升最高 6 倍,指令遵循提升 3 倍,并修复了推理循环问题。团队将这种方法称为 Weight-Space Programming,即前沿 LLM 可以原生读写神经网络权重。作者认为相比现有通过写代码启动训练任务的 RSI 流程,这提供了经由权重直接递归自我改进的更原生路径。
Current read
有人让 GPT-6 Astra 直接改 Qwen3.5-0.8B 的权重,不动一行训练代码,数学就快了 6 倍,这思路挺新鲜。
11 sources31° AI heatLast updated 2026/11/2 20:31:05
Evidence
11 sourcesRelated source 1李继刚
View sourceRelated source 2Jerry Liu
View sourceRelated source 3Thomas Wolf
View sourceRelated source 4The Rundown AI
View sourceRelated source 5orange.ai
View sourceRelated source 6shao__meng
View sourceRelated source 7歸藏(guizang.ai)
View sourceRelated source 8IT之家
View sourceRelated source 9kimmonismus
View sourceRelated source 10rohanpaul_ai
View sourceConflict check
Dedup data only shows that these sources discuss the same event, not that they agree. There is no structured supporting or refuting evidence yet, so no conflict is inferred automatically.