Research map / Probabilistic and scientific modeling / Molecules and proteins
Mitigating Reward Hacking in RLHF via Bayesian Non-negative Reward Modeling
Zhibin Duan, Guowei Rong, Zhuo Li, Bo Chen, Mingyuan Zhou, Dandan Guo
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 |
|---|---|---|---|
| add Bayesian non-negative factor reward model (BNRM) on top of BT | BT reward model (Gemma-2B-it, 40K UF) · Unified-Feedback val | accuracy: 68.8 -> 74.2 | Improved |
| add BNRM on top of GRM-SFT | GRM-SFT (Gemma-2B-it, 40K UF) · HHH-Alignment | accuracy: 78.7 -> 82.4 | Improved |
| add BNRM on top of BT | BT reward model · RewardBench Safety (Gemma-2B-it, 40K UF) | accuracy: 95.8 -> 95.6 | Mixed |
| add BNRM to the already strong Skywork reward model | Skywork-Reward-Llama-3.1-8B · RewardBench average | score: 93.1 -> 93.6 | Improved |
| use BNRM reward model for PPO | Llama3.1-8B-Instruct policy · 8-benchmark average accuracy | accuracy: 62.83 -> 74.98 | Improved |
| use BNRM reward model for PPO | OpenRLHF-Llama3-8B-SFT policy · 8-benchmark average accuracy | accuracy: 55.68 -> 62.25 | Improved |
| use BNRM reward model for PPO | OpenRLHF-Llama3-8B-SFT policy · Hellaswag | accuracy: 72.51 -> 60.68 | Hurt |
| use BNRM reward model for PPO | Llama3.1-8B-Instruct policy · GSM8K 4-shot, Hellaswag, MMLU | accuracy: 83.93 -> 82.49 (GSM8K), 77.21 -> 74.66 (Hellaswag), 72.31 -> 70.72 (MMLU) | Hurt |
| use BNRM-aligned policy versus the base policy | Llama-3.1-8B-Instruct · Arena-Hard-v0.1 (GPT-4.1 judge) | win rate: 0.4989 win vs 0.2239 lose | Improved |
| replace BT with BNRM in low-resource training | BT (Gemma-2B-it) · RewardBench, 1K UF samples | accuracy | Improved |
| replace BT with BNRM under label noise | BT (Gemma-2B-it, 40K UF) · noise ratio 0.4, OOD benchmarks | accuracy: up to +16.7 | Improved |
| use BNRM instead of BT for length debiasing | BT (Gemma-2B-it) · RM-Bench Hard, length-reward Pearson r | Pearson correlation: 0.488 -> 0.123 | Improved |
Closest papers by meaning
- Bayesian Preference Learning for Test-Time Steerable Reward Models (NeurIPS 2026 · similarity 0.90)
- InfoRM: Mitigating Reward Hacking in RLHF via Information-Theoretic Reward Modeling (NeurIPS 2024 · similarity 0.90)
- RRM: Robust Reward Model Training Mitigates Reward Hacking (ICLR 2025 · similarity 0.89)
- WARM: On the Benefits of Weight Averaged Reward Models (ICML 2024 · similarity 0.89)
- On the Robustness of Reward Models for Language Model Alignment (ICML 2025 · similarity 0.89)
- Regularizing Hidden States Enables Learning Generalizable Reward Model for LLMs (NeurIPS 2024 · similarity 0.88)
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.
