Event ·Multi-source

Qwen 3.8 27B 在笔记本上也能直接读懂模型权重

Rikiya Takehi、Jayden Teoh、Alex Zhang 和 Omar Khattab 发布论文《Sparks of Weight-Space Programming》,提出 Weight-Space Programming 概念:GPT-6 Astra 仅凭原始权重就能判断网络训练目标,并在无梯度情况下编辑 Qwen3.5-0.8B 的 0.016% 权重,将 MATH-500 从 7.9% 提升到 48.1%,IFEval 提升 3 倍。Simon Willison 复现了论文中的游戏识别实验,发现 Qwen 3.8 27B 在他的笔记本上无需任何工具也能从权重中判断网络训练的是 Breakout、Pong、Pac-Man、Freeway 四款 10x10 网格游戏中的哪一款。Matthew Honnibal 和 Sean Goedecke 则批评论文混淆了有工具和无工具两类任务,指出 Qwen3.5-0.8B 修复例子本质是人类 ML 研究员的常规操作。

Current read

一篇说 LLM 能直接读写神经网络权重的论文,Simon Willison 亲测 Qwen 3.8 27B 笔记本上就复现了,还挺有意思。

11 sources55° AI heatLast updated 2026/11/3 16:29:48

Evidence

11 sources
Primary sourceSimon Willison’s Weblog
View source
Related source 1Thomas Wolf
View source
Related source 3shao__meng
View source
Related source 7The Rundown AI
View source
Related source 9歸藏(guizang.ai)
View source

Conflict 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.

Qwen 3.8 27B can interpret model weights, matching paper claims · 事件专题 | AITOP