Features as Rewards

goodfire.ai · 1 month ago · 36

Anthropic researchers developed RLFR (Reinforcement Learning from Feature Rewards), a method that uses lightweight probes on a model's internal representations as reward signals to reduce hallucinations in language models. The approach reduced hallucinations in Gemma-3-12B-IT by 58% at approximately 90 times lower cost than using an LLM-as-judge alternative, while maintaining the ability to monitor and intervene at test time. The method enables more efficient training for open-ended tasks where ground truth verification is expensive, by leveraging factual information already present in the model's internal activations.