Research map / Large language models / Alignment and preferences
Multiplayer Nash Preference Optimization
Fang Wu, Xu Huang, Weihao Xuan, Zhiwei Zhang, Yijia Xiao, Guancheng Wan, Xiaomin Li, Bing Hu, Peng Xia, Jure Leskovec, Yejin Choi
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 |
|---|---|---|---|
| increase number of players in TD-MNPO from 1 to 3 (ablation) | TD-MNPO · AlpacaEval 2.0 (LC WR) | LC win rate: 53.32 -> 57.27 | Improved |
| increase number of players from 3 to 4 (ablation) | TD-MNPO · AlpacaEval 2.0 (LC WR) | LC win rate: 57.27 -> 57.42 | No clear effect |
| use 3 players instead of 2 in HT-MNPO (Athene-RM-8B) (ablation) | HT-MNPO · AlpacaEval 2.0 (LC WR) | LC win rate: 55.71 -> 59.64 | Improved |
| use 3 players instead of 2 in HT-MNPO (ArmoRM-Llama3) (ablation) | HT-MNPO · AlpacaEval 2.0 (LC WR) | LC win rate: 55.43 -> 57.63 | Improved |
| use 3 players instead of 2 in HT-MNPO (Skywork-Reward-V2) (ablation) | HT-MNPO · AlpacaEval 2.0 (LC WR) | LC win rate: 54.25 -> 56.01 | Improved |
| replace INPO with TD-MNPO on Llama-3-8B-it (ablation) | Llama-3-8B-it · AlpacaEval 2.0 (LC WR) | LC win rate: 41.48 -> 42.94 | Improved |
| replace INPO with TD-MNPO on Gemma-2-9B-it | Gemma-2-9B-it · Arena-Hard (WR) | win rate: 48.03 -> 52.26 | Improved |
| replace INPO with TD-MNPO on Gemma-2-9B-it | Gemma-2-9B-it · AlpacaEval 2.0 (LC WR) | LC win rate: 56.09 -> 57.27 | Improved |
| use TD-MNPO instead of INPO under different judges (GPT-4-1106-preview) | Gemma-2-9B-it · AlpacaEval 2.0 (3 runs) | LC win rate: 49.30 -> 54.05 | Improved |
| use TD-MNPO instead of baselines on academic benchmarks | Gemma-2-9B-it · 7 instruction/knowledge/commonsense benchmarks | average accuracy: 70.28 -> 71.08 | Improved |
| use TD-MNPO instead of SimPO regarding TruthfulQA degradation | Gemma-2-9B-it · TruthfulQA | accuracy: 63.40 -> 70.26 | Improved |
| use TD-MNPO instead of baselines on math/coding | Gemma-2-9B-it · Math and coding average | average score: 47.33 -> 48.10 | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- SimPO — SimPO: Simple Preference Optimization with a Reference-Free Reward (NeurIPS 2024) · better on 1 of 1 shared result
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- Iterative Nash Policy Optimization: Aligning LLMs with General Preferences via No-Regret Learning (ICLR 2025 · similarity 0.95)
- Nash Learning from Human Feedback (ICML 2024 · similarity 0.93)
- Self-Play Preference Optimization for Language Model Alignment (ICLR 2025 · similarity 0.92)
- Magnetic Preference Optimization: Achieving Last-iterate Convergence for Language Model Alignment (ICLR 2025 · similarity 0.90)
- Robust Nash Alignment under Preference Uncertainty (NeurIPS 2026 · similarity 0.90)
- A Minimaximalist Approach to Reinforcement Learning from Human Feedback (ICML 2024 · 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.
