RLBFF: Binary Flexible Feedback to bridge between Human Feedback & Verifiable Rewards

In three linesRLBFF combines human feedback and verifiable rewards for reward model training. The method extracts binary principles from natural language feedback (e.g., accuracy, code readability) and uses them as entailment tasks. Models achieve 86.2% on RM-Bench and 81.4% on JudgeBench (#1 as of September 2025). Qwen3-32B aligned with RLBFF matches o3-mini and DeepSeek R1 at <5% inference cost.

Summary generated by Claude — human-verified