Research map / Large language models / Instruction and fine-tuning
Towards Acyclic Preference Evaluation of Language Models via Multiple Evaluators
Zhengyu Hu, Jieyu Zhang, Zhihan Xiong, Alexander Ratner, Kaize Ding, Ranjay Krishna
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 graph denoising to ensemble (ablation) | PGED · response selection, HumanEval | accuracy: 69.25 -> 71.86 | Improved |
| add graph denoising to ensemble (ablation) | PGED · response selection, average of 5 benchmarks | average score: 54.87 -> 56.98 | Improved |
| add graph denoising to single evaluator Qwen2-72B (ablation) | single evaluator · response selection, HumanEval | accuracy: 60.97 -> 68.90 | Improved |
| add graph denoising to single evaluator Llama3-8B (ablation) | single evaluator · response selection, average of 5 benchmarks | average score: 52.04 -> 53.33 | Improved |
| ensemble preference graphs before denoising instead of denoising then ranking ensemble (ablation) | PGED · model ranking, AlpacaEval 2.0 | Spearman avg: 67.27 -> 72.50 | Improved |
| add graph denoising to ensemble (ablation) | PGED · model ranking, AlpacaEval 2.0 | Spearman avg: 70.72 -> 72.50 | Improved |
| add graph denoising to single evaluator Qwen2-72B (ablation) | single evaluator · model ranking, AlpacaEval 2.0 | Spearman avg: 65.04 -> 70.32 | Improved |
| combine small evaluators with PGED instead of using Qwen2-72B alone | evaluators · response selection, HumanEval | accuracy: 60.97 -> 71.86 | Improved |
| use PGED instead of Majority Voting | multiple evaluators · response selection, average of 5 benchmarks | average score: 53.73 -> 56.98 | Improved |
| list all candidates in one prompt instead of pairwise comparison (ListPreference) | Qwen2-72B · response selection, average of 5 benchmarks | average score: 53.07 -> 52.16 | Hurt |
| use PGED data selection instead of random selection | Llama-2-7B · AlpacaEval 2.0 win rate | win rate: 47.10% -> 53.50% | Improved |
| use PGED data selection instead of ContraSolver | Llama-2-7B · Koala win rate | win rate: 63.36% -> 65.91% | Improved |
Closest papers by meaning
- Peering Through Preferences: Unraveling Feedback Acquisition for Aligning Large Language Models (ICLR 2024 · similarity 0.89)
- Improving Context-Aware Preference Modeling for Language Models (NeurIPS 2024 · similarity 0.89)
- Proximalized Preference Optimization for Diverse Feedback Types: A Decomposed Perspective on DPO (NeurIPS 2025 · similarity 0.89)
- Group Preference Optimization: Few-Shot Alignment of Large Language Models (ICLR 2024 · similarity 0.88)
- Re-evaluating Open-ended Evaluation of Large Language Models (ICLR 2025 · similarity 0.88)
- SimPER: A Minimalist Approach to Preference Alignment without Hyperparameters (ICLR 2025 · 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.
