Research map / Reinforcement learning / RL from feedback
MRPO: Magnitude-Regularized Policy Optimization via L1 Constraints
Wei Han, Yuanxing Liu, Mingda Li, Ruiyu Xiao, Weinan Zhang, Ting Liu
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 the KL constraint with an L1-norm constraint in GRPO-style training | Llama3.1-8B-Instruct online preference alignment · AlpacaEval 2.0 WR | win rate (%): 53.97 -> 67.73 | Improved |
| replace the KL constraint with an L1-norm constraint | Qwen2.5-7B (RLVR) · Code-R1 pass@1 | pass@1 (%): 26.97 -> 33.56 | Improved |
| replace the KL constraint with an L1-norm constraint | Qwen2.5-7B (RLVR) · GSM8k pass@1 | pass@1 (%): 83.09 -> 84.08 | Improved |
| use L1 constraint instead of reverse KL with a weak 2B reward model | Qwen2.5-7B-Instruct · AlpacaEval 2.0 WR | win rate (%): 39.62 -> 48.51 | Improved |
| use L1 constraint instead of forward KL or JS divergence (ablation) | Llama3.1-8B-Instruct · AlpacaEval 2.0 WR | win rate (%): 50.17 -> 67.73 (DPH-JS -> MRPO-G) | Improved |
| apply L1 constraint in offline RL with binary +1/-1 rewards instead of DPO | Gemma-7b-it · offline RL average over six benchmarks | average score: 52.72 -> 53.20 | Improved |
| apply L1 constraint in offline RL | Gemma-7b-it · offline RL complex reasoning | average score: 44.61 -> 45.40 | Improved |
| apply L1 constraint in offline RL | Gemma-7b-it · offline RL general capabilities | average score: 60.82 -> 60.99 | No clear effect |
| use REINFORCE-based MRPO-R instead of GRPO-based MRPO-G (ablation) | MRPO · Arena-Hard Gemini-2 (Llama3.1-8B) | win rate (%): 81.82 -> 84.09 on Llama; 31.02 -> 29.68 on Mistral | Mixed |
| use MRPO instead of GRPO on a small 1.5B model | Qwen2.5-1.5B-Instruct · AlpacaEval 2.0 WR | win rate (%): 17.00 -> 28.46 | Improved |
| use MRPO instead of GRPO on a 3B model | Qwen2.5-3B-Instruct · AlpacaEval 2.0 WR | win rate (%): 49.10 -> 62.25 | Improved |
| use MRPO on a base model with an already high win rate on AlpacaEval 1.0 | Llama3.1-8B-Instruct · AlpacaEval 1.0 WR | win rate (%): 97.34 -> 97.59 | Improved |
Closest papers by meaning
- One-Way Policy Optimization for Self-Evolving LLMs (ICML 2026 · similarity 0.90)
- AlphaDPO: Adaptive Reward Margin for Direct Preference Optimization (ICML 2025 · similarity 0.89)
- Towards Efficient Exact Optimization of Language Model Alignment (ICML 2024 · similarity 0.89)
- Multi-Reference Preference Optimization for Large Language Models (AAAI 2025 · similarity 0.89)
- VESPO: Variational Sequence-level Soft Policy Optimization for Off-Policy LLM Training (NeurIPS 2026 · similarity 0.89)
- Explicit Preference Optimization: No Need for an Implicit Reward Model (ICML 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.
