Research map / Large language models / Alignment and preferences
RSPO: Regularized Self-Play Alignment of Large Language Models
Xiaohang Tang, Sangwoong Yoon, Seongho Son, Huizhuo Yuan, Quanquan Gu, Ilija Bogunovic
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 reverse KL regularization to SPPO loss (ablation) | SPPO · AlpacaEval 2.0, Mistral-7B, iterations | LC win rate: 28.53 -> 34.21 (Iter3) | Improved |
| add IS-forward plus reverse KL regularization (ablation) | SPPO · AlpacaEval 2.0, Mistral-7B | LC win rate: 28.53 -> 35.44 | Improved |
| add chi-square regularization (ablation) | SPPO · AlpacaEval 2.0, Mistral-7B, Iter3 | LC win rate: 28.53 -> 29.97 | Improved |
| increase forward KL regularization temperature (ablation) | RSPO (IS-Forward KL) · Mistral-7B | response length and win rate | Hurt |
| increase reverse KL regularization temperature (ablation) | RSPO (Reverse KL) · Mistral-7B | response length | Improved |
| add reverse KL regularization (ablation) | SPPO · Self-BLEU diversity, Mistral-7B Iter3 | Self-BLEU: 0.758 -> 0.691 | Improved |
| run SPPO for more iterations (up to 9) (ablation) | SPPO · AlpacaEval 2.0, Mistral-7B | LC win rate: 28.53 -> 29.17 | No clear effect |
| use stronger preference model GPM-8B instead of PairRM (ablation) | SPPO / RSPO on LLaMA-8B · AlpacaEval 2.0 | LC win rate: +0.44 | Improved |
| add RSPO regularization on LLaMA-8B | SPPO · AlpacaEval 2.0 | LC win rate: 38.77 -> 43.66 | Improved |
| add RSPO regularization on Gemma-2B | SPPO · AlpacaEval 2.0 | LC win rate: 50.54 -> 51.83 | Improved |
Later papers that compared with this method
1 later accepted paper reports RSPO as a baseline in their results tables; 1 of them report a better number on most of the results they share with it.
- PADD: Path-Aligned Decompression Distillation for Non-Router Teacher to Guide MoE Student Learning (ICML 2026 · better on 7 of 7 shared results)
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 2 of 2 shared results
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
- Self-Play Preference Optimization for Language Model Alignment (ICLR 2025 · similarity 0.93)
- Magnetic Preference Optimization: Achieving Last-iterate Convergence for Language Model Alignment (ICLR 2025 · similarity 0.92)
- Iterative Nash Policy Optimization: Aligning LLMs with General Preferences via No-Regret Learning (ICLR 2025 · similarity 0.90)
- AMPO: Active Multi Preference Optimization for Self-play Preference Selection (ICML 2025 · similarity 0.90)
- Stackelberg Self-Annotation: A Robust Approach to Data-Efficient LLM Alignment (NeurIPS 2025 · similarity 0.88)
- LLM Active Alignment: A Nash Equilibrium Perspective (NeurIPS 2026 · 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.
