Primary source歸藏(guizang.ai)
View sourceEvent ·Official first-hand
MIT CSAIL 用 GPT-6 Astra 直接改写 Qwen 0.8B 权重,数学能力提升 6 倍
MIT CSAIL 团队提出 Weight-Space Programming 方法,用 GPT-6 Astra 直接编辑 Qwen 0.8B 模型的权重文件。仅修改 0.016% 的权重,就让模型数学能力提升 6 倍、指令跟随能力提升 3 倍。整个过程没有梯度更新、没有重新训练或强化学习,还修复了该模型重复推理、频繁触及思考长度上限的问题。团队认为这为模型控制和可解释性提供了一套类似编程的操作体系,小模型尤其适合这种改写方式。
Current read
MIT 团队不训练不微调,直接让 GPT-6 Astra 改 Qwen 0.8B 的权重,改 0.016% 参数数学就快 6 倍,脑洞很新。
11 sources31° AI heatLast updated 2026/11/3 03:30:00
Evidence
11 sourcesRelated source 1The Rundown AI
View sourceRelated source 2shao__meng
View sourceRelated source 3Jerry Liu
View sourceRelated source 4Thomas Wolf
View sourceRelated source 5elvis
View sourceRelated source 6李继刚
View sourceRelated source 7orange.ai
View sourceRelated source 8kimmonismus
View sourceRelated source 9IT之家
View sourceRelated source 10宝玉
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.