Research map / Large language models / Alignment and preferences
RLBFF: Binary Flexible Feedback to bridge between Human Feedback & Verifiable Rewards
Zhilin Wang, Jiaqi Zeng, Olivier Delalleau, Ellie Evans, Daniel Egert, Hoo-Chang Shin, Felipe Soares, Yi Dong, Oleksii Kuchaiev
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| train with principle-conditioned Yes/No (Flexible Principles) instead of Bradley-Terry | Llama-3.3-70B scalar RM · RM-Bench | overall accuracy: 78.5 -> 83.6 | Improved |
| train with principle-conditioned Yes/No (Flexible Principles) instead of Bradley-Terry | Llama-3.3-70B scalar RM · JudgeBench | overall accuracy: 68.9 -> 76.3 | Improved |
| train with principle-conditioned Yes/No instead of Bradley-Terry | Llama-3.1-8B scalar RM · JudgeBench (8B base) | overall accuracy: 62.9 -> 64.0 | Improved |
| replace scalar RM with reasoning GenRM trained with GRPO | Flexible Principles RM · JudgeBench | overall accuracy: 76.3 -> 81.4 | Improved |
| replace scalar RM with reasoning GenRM | Flexible Principles RM · PrincipleBench | overall micro-average: 91.6 -> 83.8 | Hurt |
| raise consensus similarity threshold from 0.8 to 0.9 (ablation) | Flexible Principles ScalarRM · JudgeBench | overall accuracy: 76.3 -> 73.7 | Hurt |
| lower consensus similarity threshold from 0.8 to 0.7 (ablation) | Flexible Principles ScalarRM · JudgeBench | overall accuracy: 76.3 -> 72.3 | Hurt |
| train on multiple flexible principles instead of one fixed principle (test-time fixed) (ablation) | Flexible Principles ScalarRM · RM-Bench | overall accuracy: 79.9 -> 81.9 | Improved |
| allow flexible principle selection at test time instead of fixing accuracy of information (ablation) | Flexible Principles ScalarRM · JudgeBench | overall accuracy: 70.9 -> 76.3 | Improved |
| remove equal Yes/No weighting (keep natural imbalance) (ablation) | Flexible Principles ScalarRM · RM-Bench | overall accuracy: 81.5 -> 83.6 | No clear effect |
| train Qwen3-32B with RLBFF GenRM reward instead of Bradley-Terry RM | Qwen3-32B · Arena Hard v2 | score: 47.5 -> 55.6 | Improved |
| train Qwen3-32B with RLBFF GenRM reward instead of Bradley-Terry RM | Qwen3-32B · WildBench Overall | score: 67.38 -> 70.33 | Improved |
Closest papers by meaning
- Uni-RLHF: Universal Platform and Benchmark Suite for Reinforcement Learning with Diverse Human Feedback (ICLR 2024 · similarity 0.91)
- Exploring Data Scaling Trends and Effects in Reinforcement Learning from Human Feedback (NeurIPS 2025 · similarity 0.91)
- Zeroth-Order Policy Gradient for Reinforcement Learning from Human Feedback without Reward Inference (ICLR 2025 · similarity 0.91)
- Provably Efficient Online RLHF with One-Pass Reward Modeling (NeurIPS 2025 · similarity 0.91)
- A Theoretical Framework for Partially-Observed Reward States in RLHF (ICLR 2025 · similarity 0.90)
- The Surprising Effectiveness of Negative Reinforcement in LLM Reasoning (NeurIPS 2025 · similarity 0.90)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
