Features as Rewards

This viewer accompanies our blog post on reducing hallucinations via RLFR. It shows rollouts from Gemma 3 12B, before (base) and after RLFR (policy), on selected items from LongFact++, a benchmark for open-ended, long-form factuality in LLM responses.

Highlights show detected hallucinations and interventions. Click any highlight to open intervention details in a side drawer.

Rollouts with a on their pill have author commentary.