Research map / Large language models / Alignment and preferences
Mitigating Mismatch within Reference-based Preference Optimization
Suqin Yuan, Xingrui Yu, Jiyang Zheng, Lei Feng, Dadong Wang, Ivor Tsang, Tongliang 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 |
|---|---|---|---|
| clip the reference margin at zero (HyPO) instead of DPO | DPO on Mistral-Base 7B · AlpacaEval 2.0 | LC win rate: 22.6 -> 32.8 | Improved |
| clip the reference margin at zero (HyPO) instead of DPO | DPO on Llama-3-Instruct 8B · Arena-Hard-v0.1 | WR: 31.5 -> 35.2 | Improved |
| clip the reference margin at zero (HyPO) instead of DPO | DPO on Mistral-Instruct 7B · AlpacaEval 2.0 | WR: 31.4 -> 47.9 | Improved |
| HyPO versus reference-free SimPO | SimPO on Llama-3-Base 8B · Arena-Hard-v0.1 | WR: 30.1 -> 33.5 | Improved |
| HyPO versus DPO averaged over four settings | DPO · AlpacaEval 2.0 and Arena-Hard | relative improvement: 41.2% average relative improvement | Improved |
| HyPO on different data and models | DPO · HelpSteer2 | LC win rate: 16.2 -> 22.3 | Improved |
| HyPO on larger model | DPO · Mistral-Nemo-Instruct 12B | LC win rate: 50.4 -> 55.7 | Improved |
| HyPO on Qwen-2.5-7B | DPO · Qwen-2.5-7B | LC win rate: 27.9 -> 38.0 | Improved |
| add SFT loss on chosen response to DPO (ablation) | DPO · AlpacaEval 2.0 (Llama-3-Base 8B) | LC win rate: 24.3 -> 26.1 | Improved |
| add SFT loss on chosen response to DPO (ablation) | DPO · AlpacaEval 2.0 (Llama-3-Base 8B) | raw win rate: 21.9 -> 21.9 | No clear effect |
| HyPO instead of DPO + SFT-on-chosen | DPO + SFT-on-chosen · AlpacaEval 2.0 (Llama-3-Base 8B) | LC win rate: 26.1 -> 28.8 | Improved |
| HyPO instead of DPO on agreement rate | DPO · Llama-3-Base 8B | agreement rate: 55.9 -> 60.4 | 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 12 of 12 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
- Proximalized Preference Optimization for Diverse Feedback Types: A Decomposed Perspective on DPO (NeurIPS 2025 · similarity 0.92)
- KL Penalty Control via Perturbation for Direct Preference Optimization (NeurIPS 2025 · similarity 0.91)
- Multi-Reference Preference Optimization for Large Language Models (AAAI 2025 · similarity 0.91)
- 3D-Properties: Identifying Challenges in DPO and Charting a Path Forward (ICLR 2025 · similarity 0.91)
- Towards Robust Alignment of Language Models: Distributionally Robustifying Direct Preference Optimization (ICLR 2025 · similarity 0.90)
- Less is More: Improving LLM Alignment via Preference Data Selection (NeurIPS 2025 · 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.
