跳到正文
Goodfire Research·· 2026-02-11AI 评分63

Goodfire 发布 RLFR 方法:用内部特征探针作奖励将 Gemma-3-12B-IT 幻觉率降低 58%

Features as Rewards: Using Interpretability to Reduce Hallucinations

AI 导读

Goodfire Research 发布论文,提出 RLFR(Reinforcement Learning from Feature Rewards),用模型内部激活上的轻量探针作为 RL 奖励信号来减少幻觉。

来源:Goodfire Research · goodfire.com