Research map / Large language models / Alignment and preferences
The Hidden Link between RLHF and Contrastive Learning
Xufei Lv, Kehai Chen, Haoyuan Sun, Xuefeng Bai, Min Zhang, Houde 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 DV/MINE bound (DPO) with JS estimator (MIO) | Mistral-7B-SFT · Arena-Hard | score: 16.3 -> 23.6 | Improved |
| replace DPO with MIO | Mistral-7B-SFT · AlpacaEval 2 | win rate: 16.3 -> 21.2 | Improved |
| replace DPO with MIO | Mistral-7B-SFT · MT-Bench | score: 7.3 -> 7.3 | No clear effect |
| replace DPO with MIO | Qwen2.5-7B-Instruct · GSM8K | accuracy: 73.46 -> 75.39 | Improved |
| replace DPO with MIO | LLaMA3-8B-SFT · MUSR | accuracy: 39.11 -> 42.81 | Improved |
| replace DPO with MIO on rare-rejected examples | Qwen2.5-7B-Instruct · Local update diagnostic, buckets 0-6 | change in chosen likelihood: DPO -0.1825 vs MIO 3.3761 (bucket 0) | Improved |
| replace DPO with MIO in toy model | toy MLP policy · toy model, rejected prob near zero | chosen/rejected likelihood dynamics | Improved |
| use MIO instead of DPO in online preference optimization | Qwen3.5-9B · TL;DR online | win rate: 30.1 -> 35.5 | Improved |
| replace DPO with MIO | GPT-J SFT · CNN/DailyMail OOD (temp 0) | GPT-4 win rate: 0.36 -> 0.38 | Improved |
| increase beta from 2.0 to 20.0 (ablation) | MIO, Mistral-7B-SFT · IFEval | score: 46.37 -> 36.59 | Hurt |
| vary beta (ablation) | MIO, Mistral-7B-SFT · GSM8K | accuracy: 39.91 (beta 2.0), 40.16 (beta 10.0), 38.74 (beta 20.0) | No clear effect |
| add 10% label noise to training data (ablation) | MIO vs DPO, Qwen2.5-7B-Instruct · MUSR | accuracy: DPO 38.5 -> 36.7, MIO 38.5 -> 36.0 | Hurt |
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
- Value-Incentivized Preference Optimization: A Unified Approach to Online and Offline RLHF (ICLR 2025 · similarity 0.89)
- Online Preference Alignment for Language Models via Count-based Exploration (ICLR 2025 · similarity 0.89)
- Generative RLHF-V: Learning Principles from Multi-modal Human Preference (NeurIPS 2025 · similarity 0.88)
- MaxMin-RLHF: Alignment with Diverse Human Preferences (ICML 2024 · similarity 0.88)
- Beyond Reverse KL: Generalizing Direct Preference Optimization with Diverse Divergence Constraints (ICLR 2024 · similarity 0.88)
- Iterative Preference Learning from Human Feedback: Bridging Theory and Practice for RLHF under KL-constraint (ICML 2024 · 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.
