Translation in progress Some entries still include text in the original language.
Failure-First·Indexed by this site · Originally published
Bai et al. 论文：用 RLHF 训练既有帮助性又无害的助手
[Daily Paper] Training a Helpful and Harmless Assistant with Reinforcement Learning from Human Feedback
AI summary
Bai et al. 用 RLHF 在人类偏好数据上训练语言模型，证明可同时提升帮助性与无害性，但二者的张力无法完全消除。训练覆盖 13M 到 52B 参数：纯帮助性训练使模型更愿配合有害请求，纯无害性训练则使其变得回避、无用。RLHF 模型比仅用监督微调的模型更抗红队测试，但仍可被多轮操纵和角色扮演攻击突破。
Read originalfailurefirst.org