Translation in progress Some entries still include text in the original language.
Source OpenAI Alignment Research
Tomek Korbak, Cameron Raymond, Micah Carroll, Marcus Williams, Mikita Balesni, Alan Guo, Jason Wolfe, Akshay Jagadeesh, Ian Kivlichan
Indexed by this site 2026-09-30 22:18 12 days ago
Originally published Mar 28
Selected 2026-09-30 22:33
Category Alignment OpenAI Alignment Research · Tomek Korbak, Cameron Raymond, Micah Carroll, Marcus Williams, Mikita Balesni, Alan Guo, Jason Wolfe, Akshay Jagadeesh, Ian Kivlichan · Indexed by this site 2026-09-30 22:18 · Originally publishedMar 28 Attention50
OpenAI 研究：对齐中间训练能泛化多远 How far does alignment midtraining generalize?
AI summary
OpenAI 对齐研究团队在类似 o4-mini 的模型上复现并扩展了 Tice 等人的对齐中间训练实验，用 230k 篇文档（约 340M token）分别做对齐与不对齐中间训练，再接入推理后训练。
Why it was selected
OpenAI 公开了一组对齐中间训练的负面结果，说明该干预在推理后训练和真实智能体场景中难以泛化，为后续对齐先验研究提供了可比较的基线。
Read original alignment.openai.com Why it was selected Attention50
Attention 0–100: content value contributes up to 70 and circulation up to 30; circulation decays over time. How it was rated
OpenAI 公开了一组对齐中间训练的负面结果，说明该干预在推理后训练和真实智能体场景中难以泛化，为后续对齐先验研究提供了可比较的基线。