Event ·Multi-source

OpenAI研究：RL训练美德，好行为泛化至44项评测

OpenAI发现对齐大模型时存在涌现失调现象，即坏行为会泛化。他们反向实验用RL训练模型诚实、谦逊、可纠正等特质，仅混入小部分此类数据。结果在训练领域内模型变得更诚实透明；在44个独立评测（未见过）中，欺骗、谄媚、有害建议等行为全面下降，即使只用健康数据训练，非健康领域也受益。模型在对抗性提示和恶意微调下更坚韧，正常指令仍可听从。

Current read

OpenAI这篇论文反直觉：用RL教模型做好事，坏行为自己就减少了。实验覆盖44个新场景，效果还抗攻击。值得一看。

11 sources86° AI heatLast updated 2026/6/18 22:40:08

Evidence

11 sources

Primary sourceorange.ai

2026/6/18 22:40:08View source

Related source 1OpenAI

2026/6/18 21:34:40View source

Related source 2Decoder

2026/6/19 10:08:27View source

Related source 3berryxia

2026/6/20 17:50:33View source

Related source 4shao__meng

2026/6/17 00:53:38View source

Related source 5IT之家

2026/6/17 02:06:53View source

Related source 6marktechpost

2026/6/17 05:49:54View source

Related source 7arXiv: OpenAI

2026/6/17 08:04:12View source

Related source 8AI Will

2026/6/17 09:19:40View source

Related source 9小互

2026/6/17 13:54:45View source

Related source 10Aadit Sheth

2026/6/17 19:22:57View 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.

Open the story