Research map / Large language models / Alignment and preferences
Unbiased Alignment for Large Language Models with Noisy Preferences
Jialiang Wang, Xianming Liu, Xiong Zhou, Hui Liu, Haoliang Li
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 DPO loss with UDPO loss | DPO, Llama-3.2-3B · HH+TL;DR average | win rate vs SFT (%): 71.4 -> 85.9 | Improved |
| replace DPO loss with UDPO loss | DPO, Qwen-3-1.7B · HH+TL;DR average | win rate vs SFT (%): 65.9 -> 79.4 | Improved |
| replace RM loss with URM loss | RM, Qwen-3-1.7B · HH+TL;DR average | win rate vs SFT (%): 63.9 -> 70.1 | Improved |
| replace RM loss with URM loss | RM, Llama-3.2-3B · HH+TL;DR average | win rate vs SFT (%): 75.5 -> 79.4 | Improved |
| replace RM loss with URM loss | RM, Qwen-3-8B · UltraFeedback | win rate vs SFT (%): 54.3 -> 64.6 | Improved |
| replace DPO loss with UDPO loss | DPO, Llama-3.1-8B · UltraFeedback | win rate vs SFT (%): 65.4 -> 74.3 | Improved |
| replace DPO loss with UDPO loss | DPO, Qwen-3-8B · UltraFeedback | win rate vs SFT (%): 56.3 -> 60.6 | Improved |
| add label smoothing (cRM, cDPO) to the loss | RM and DPO, Llama-3.2-3B · HH+TL;DR average | win rate vs SFT (%): 75.5 -> 75.7 (RM); 71.4 -> 70.7 (DPO) | No clear effect |
| add label smoothing to DPO (cDPO) | DPO, Qwen-3-1.7B · HH+TL;DR average | win rate vs SFT (%): 65.9 -> 63.6 | Hurt |
| use URM instead of rRM under 40% flipped labels | rRM, Llama-3.2-3B · TL;DR 40% flip | win rate vs SFT (%): 68.6 -> 77.5 | Improved |
| use UDPO instead of rDPO under instance-dependent noise 40% | rDPO, Qwen-3-1.7B · HH IDN 40% | win rate vs SFT (%): 63.7 -> 68.0 | Improved |
| use URM instead of RM under instance-dependent noise 40% | RM, Qwen-3-1.7B · HH IDN 40% | win rate vs SFT (%): 46.0 -> 54.6 | Improved |
Closest papers by meaning
- Noise Contrastive Alignment of Language Models with Explicit Rewards (NeurIPS 2024 · similarity 0.92)
- Towards Robust Alignment of Language Models: Distributionally Robustifying Direct Preference Optimization (ICLR 2025 · similarity 0.92)
- Doubly Robust Alignment for Large Language Models (NeurIPS 2025 · similarity 0.92)
- Cal-DPO: Calibrated Direct Preference Optimization for Language Model Alignment (NeurIPS 2024 · similarity 0.92)
- ComPO: Preference Alignment via Comparison Oracles (NeurIPS 2025 · similarity 0.92)
- Towards Efficient Exact Optimization of Language Model Alignment (ICML 2024 · 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.
