Primary sourceSuperTechFans
View sourceEvent ·Multi-source
MIT CSAIL 用 Astra 编辑 Qwen3.5-0.8B 权重,MATH-500 从 7.9% 升至 48.1%
MIT CSAIL 团队让通用语言模型 Astra 通过 Python 分析和最多 50 批练习反馈,修改 Qwen3.5-0.8B 两层中共 131072 个权重,约占总参数 0.016%。修改后 MATH-500 从 7.9% 升到 48.1%,GSM8K 从 13.8% 升到 67.1%,IFEval 从 21.0% 升到 62.3%。主要增益来自解除推理循环,MATH-500 触发输出上限的比例从 93% 降到 33%。honnibal.dev 的反驳指出,直接读权重只在小网络实验中出现,Qwen 结果依赖工具和反馈,二者不能当作同一种能力。
Current read
MIT 团队不改训练直接改 Qwen 小模型的 13 万个权重,数学分数翻了六倍,还有一篇逐条反驳,正反两篇都值得看。
11 sources81° AI heatLast updated 2026/11/3 23:47:22
Evidence
11 sourcesRelated source 1shao__meng
View sourceRelated source 2IT之家
View sourceRelated source 3rohanpaul_ai
View sourceRelated source 4The Rundown AI
View sourceRelated source 5AIGCLINK
View sourceRelated source 6Simon Willison’s Weblog
View sourceRelated source 7歸藏(guizang.ai)
View sourceRelated source 8arXiv cs.LG
View sourceRelated source 9宝玉
View sourceRelated source 10Gorden Sun
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.