Research map / Large language models / Alignment and preferences
Reliability-Aware LLM Alignment from Inconsistent Human Feedback
Jingyi Huang, Ruohan Zong, Yujun Feng, Liran Ma, Lanyu Shang, Yang Zhang
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 RGPO | DPO · MultiPref (Llama-3-8B-Instruct) | Arena-Hard win rate: 43.70% to 46.60% | Improved |
| add RGPO | IPO · MultiPref (Llama-3-8B-Instruct) | Arena-Hard win rate: 51.90% to 53.10% | Improved |
| add RGPO | SimPO · MultiPref (Qwen2.5-7B-Instruct) | AlpacaEval 2 raw win rate: 69.07 -> 85.36 | Improved |
| add RGPO | SimPO · HelpSteer2-Disagreement (Llama-3-8B-Instruct) | AlpacaEval 2 raw win rate: 12.31 -> 25.60 | Improved |
| add RGPO | DPO · HelpSteer2-Disagreement (Qwen2.5-7B-Instruct) | AlpacaEval 2 raw win rate: 40.31 -> 39.83 | No clear effect |
| add RGPO | DPO · MultiPref (Qwen2.5-7B-Instruct) | Arena-Hard win rate: 78.20 -> 77.10 | Hurt |
| add ILRE (iterative latent reliability estimation) (ablation) | RGPO-DPO · MultiPref (Llama-3-8B-Instruct) | Arena-Hard win rate: 42.50 -> 46.60 | Improved |
| add RACO (reliability-aware consistency optimization) (ablation) | RGPO-DPO · MultiPref (Llama-3-8B-Instruct) | Arena-Hard win rate: 45.40 -> 46.60 | Improved |
| replace confidence filtering with RGPO (ablation) | DPO · MultiPref (Llama-3-8B-Instruct) | AlpacaEval 2 raw win rate: 38.67 -> 40.74 | Improved |
| add confidence filtering (ablation) | DPO · MultiPref (Llama-3-8B-Instruct) | AlpacaEval 2 LC win rate: 35.55 -> 36.08 | Mixed |
| add Tanh scaling of consistency weight (ablation) | RGPO-DPO · MultiPref (Llama-3-8B-Instruct) | AlpacaEval 2 LC win rate: 35.54 -> 38.30 | Improved |
| RGPO ILRE yields more tie labels than uniform ensemble (ablation) | preference construction · MultiPref | tie fraction: 12.27% to 29.03% | 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 5 of 5 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
- RLTHF: Targeted Human Feedback for LLM Alignment (ICML 2025 · similarity 0.93)
- Doubly Robust Alignment for Large Language Models (NeurIPS 2025 · similarity 0.92)
- Preference Ranking Optimization for Human Alignment (AAAI 2024 · similarity 0.92)
- Towards Efficient Exact Optimization of Language Model Alignment (ICML 2024 · similarity 0.92)
- Robust LLM Alignment via Distributionally Robust Direct Preference Optimization (NeurIPS 2025 · similarity 0.91)
- Online Preference Alignment for Language Models via Count-based Exploration (ICLR 2025 · similarity 0.91)
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.
