AI Paper Atlas

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

ICLR 2026 · Large language models › Alignment and preferences · approach: DPO design choices · 1 citations (OpenAlex)

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 changedWhereMeasuredEffect
clip the reference margin at zero (HyPO) instead of DPODPO on Mistral-Base 7B · AlpacaEval 2.0LC win rate: 22.6 -> 32.8Improved
clip the reference margin at zero (HyPO) instead of DPODPO on Llama-3-Instruct 8B · Arena-Hard-v0.1WR: 31.5 -> 35.2Improved
clip the reference margin at zero (HyPO) instead of DPODPO on Mistral-Instruct 7B · AlpacaEval 2.0WR: 31.4 -> 47.9Improved
HyPO versus reference-free SimPOSimPO on Llama-3-Base 8B · Arena-Hard-v0.1WR: 30.1 -> 33.5Improved
HyPO versus DPO averaged over four settingsDPO · AlpacaEval 2.0 and Arena-Hardrelative improvement: 41.2% average relative improvementImproved
HyPO on different data and modelsDPO · HelpSteer2LC win rate: 16.2 -> 22.3Improved
HyPO on larger modelDPO · Mistral-Nemo-Instruct 12BLC win rate: 50.4 -> 55.7Improved
HyPO on Qwen-2.5-7BDPO · Qwen-2.5-7BLC win rate: 27.9 -> 38.0Improved
add SFT loss on chosen response to DPO (ablation)DPO · AlpacaEval 2.0 (Llama-3-Base 8B)LC win rate: 24.3 -> 26.1Improved
add SFT loss on chosen response to DPO (ablation)DPO · AlpacaEval 2.0 (Llama-3-Base 8B)raw win rate: 21.9 -> 21.9No clear effect
HyPO instead of DPO + SFT-on-chosenDPO + SFT-on-chosen · AlpacaEval 2.0 (Llama-3-Base 8B)LC win rate: 26.1 -> 28.8Improved
HyPO instead of DPO on agreement rateDPO · Llama-3-Base 8Bagreement rate: 55.9 -> 60.4Improved

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.

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

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.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy