Research map / Reinforcement learning / RL from feedback
URPO: A Unified Reward & Policy Optimization Framework for Large Language Models
Songshuo Lu, Hua Wang, Zhi Chen, Yaohua Tang
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 separate RM-gen reward with unified URPO | GRPO on Qwen2.5-7B Base · AlpacaEval 2.0 | win rate: 42.24 -> 44.84 | Improved |
| replace separate RM-gen reward with unified URPO | GRPO on Qwen2.5-7B Base · AIME 24&25 (mean@8) | accuracy: 12.08 -> 16.46 | Improved |
| replace separate RM-gen reward with unified URPO | GRPO on Qwen2.5-7B+SFT · Math average (SFT start) | average accuracy: 31.64 -> 31.83 | Improved |
| replace separate RM-gen reward with unified URPO | GRPO on Qwen2.5-7B+SFT · GSM8K (SFT start) | accuracy: 90.37 -> 87.79 | Hurt |
| replace RM-gen reward with URPO | GRPO on Qwen3-8B-Base · Math average | average accuracy: 39.75 -> 39.57 | No clear effect |
| replace RM-gen reward with URPO | GRPO on OctoThinker Llama3.1-8B-Base · AlpacaEval 2.0 | win rate: 36.15 -> 41.49 | Improved |
| add instruction data to the P+R mix (1:1:0 -> 1:1:1) (ablation) | URPO · AlpacaEval 2.0 | win rate: 31.43 -> 44.84 | Improved |
| add reasoning data to the P+I mix (1:0:1 -> 1:1:1) (ablation) | URPO · Math average | average accuracy: 28.58 -> 35.66 | Improved |
| add reasoning data to preference data (ablation) | URPO · RewardBench | mean score: 83.55 -> 85.15 | Improved |
| add preference data to the R+I mix (0:1:1 -> 1:1:1 style) (ablation) | URPO · RewardBench | mean score: 62.39 -> 85.15 | Improved |
| double preference ratio (1:1:0 -> 2:1:0) (ablation) | URPO · RewardBench | mean score: 85.15 -> 85.31 | No clear effect |
| raise preference ratio (1:1:1 -> 3:1:1) (ablation) | URPO · Math average / AlpacaEval | metric: Math 35.66 -> 35.77; AlpacaEval 44.84 -> 42.11 | Mixed |
Closest papers by meaning
- Towards Efficient Exact Optimization of Language Model Alignment (ICML 2024 · similarity 0.90)
- SR-GRPO: Stable Rank as an Intrinsic Geometric Reward for Large Language Model Alignment (NeurIPS 2026 · similarity 0.89)
- RM-Bench: Benchmarking Reward Models of Language Models with Subtlety and Style (ICLR 2025 · similarity 0.89)
- UniRRM: Unified Reasoning Reward Models Across Languages and Evaluation Paradigms (ICML 2026 · similarity 0.89)
- GVPO: Group Variance Policy Optimization for Large Language Model Post-Training (NeurIPS 2025 · similarity 0.89)
- Reducing the Probability of Undesirable Outputs in Language Models Using Probabilistic Inference (NeurIPS 2025 · similarity 0.89)
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.
