Research map / Learning theory and algorithms / Optimal transport
Optimal Transport for LLM Reward Modeling from Noisy Feedback
Licheng Pan, Haocheng Yang, Haoxuan Li, Yunsheng Lu, Yongqi Tong, Yinuo Wang, Shijian Wang, Zhixuan Chu, Lei Shen, Yuan Lu, Hao Wang
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 |
|---|---|---|---|
| replace naive pointwise training with full optimal-transport alignment using preference cost only (ablation) | reward model on noisy labels · HelpSteer (noisy labels) | MSE: 0.083 -> 0.079 | Improved |
| add semantic-embedding term to the transport cost (joint cost) under full transport (ablation) | SelectiveRM full-transport variant · HelpSteer (noisy labels) | MSE: 0.079 -> 0.067 | Improved |
| add semantic-embedding joint cost under full transport (ablation) | SelectiveRM full-transport variant · PKU-SafeRLHF (noisy labels) | R2: 0.684 -> 0.710 | Improved |
| add mass relaxation (partial transport) to preference-cost transport (ablation) | SelectiveRM preference-cost variant · HelpSteer (noisy labels) | MSE: 0.079 -> 0.065 | Improved |
| add mass relaxation (partial transport) to preference-cost transport (ablation) | SelectiveRM preference-cost variant · PKU-SafeRLHF (noisy labels) | R2: 0.684 -> 0.751 | Improved |
| add mass relaxation on top of joint-cost full transport (ablation) | SelectiveRM joint-cost variant · HelpSteer (noisy labels) | MSE: 0.067 -> 0.063 | Improved |
| add joint cost on top of partial transport (ablation) | SelectiveRM partial-transport variant · UltraFeedback (noisy labels) | MSE: 0.109 -> 0.107 | Improved |
| use SelectiveRM instead of naive training | reward model on noisy labels · PKU-SafeRLHF (noisy labels) | MSE: 0.089 -> 0.057 | Improved |
| use SelectiveRM instead of naive training | reward model on noisy labels · HelpSteer (noisy labels) | MSE: 0.083 -> 0.063 | Improved |
| use SelectiveRM instead of naive training with a larger backbone | Qwen2.5-72B reward model · PKU-SafeRLHF (backbone study) | MSE: 0.078 -> 0.056 | Improved |
| use SelectiveRM instead of naive training on LLaMA2-13B | LLaMA2-13B reward model · PKU-SafeRLHF (backbone study) | MSE: 0.088 -> 0.062 | Improved |
| use SelectiveRM reward model instead of naive for GRPO policy training | Qwen2.5-7B policy · DAN (GRPO policy safety) | safety score: 0.217 -> 0.268 | Improved |
Closest papers by meaning
- Greedy Sampling Is Provably Efficient For RLHF (NeurIPS 2025 · similarity 0.88)
- Reward Model Learning vs. Direct Policy Optimization: A Comparative Analysis of Learning from Human Preferences (ICML 2024 · similarity 0.88)
- Value-Incentivized Preference Optimization: A Unified Approach to Online and Offline RLHF (ICLR 2025 · similarity 0.88)
- Regularizing Hidden States Enables Learning Generalizable Reward Model for LLMs (NeurIPS 2024 · similarity 0.88)
- On the Robustness of Reward Models for Language Model Alignment (ICML 2025 · similarity 0.88)
- Post-hoc Reward Calibration: A Case Study on Length Bias (ICLR 2025 · 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.
