Research map / Large language models / Alignment and preferences
PiCO: Peer Review in LLMs based on Consistency Optimization
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes PiCO, an unsupervised LLM evaluation method where LLMs peer-review each other's anonymous answers and learnable reviewer confidence weights are optimized via a consistency assumption to produce a ranking closer to human crowdsourced rankings.
Problem
Evaluating LLMs without human annotations or benchmark labels, producing rankings aligned with human preference on open-ended tasks, while avoiding benchmark leakage and the cost of human evaluation.
Approach
Each LLM acts as a reviewer of anonymous answer pairs with a learnable confidence weight; response scores are weighted sums of reviewer votes, and the weights are optimized to maximize Pearson consistency between reviewer capability weights and scores, with an unsupervised elimination of weak reviewers (up to 60%).
Key results
- PiCO reaches Spearman 0.89 on MT-Bench and Chatbot Arena and 0.84 on AlpacaEval with all data, versus 0.90 for Claude-3 (API) on Chatbot Arena and 0.86 for PRE on MT-Bench.
- On the 1.0-ratio setting PiCO improves permutation entropy to 0.94 on Chatbot Arena versus 1.07 for PRE, the best baseline.
- Existing benchmarks correlate weakly with human rankings: MMLU Spearman 0.53 and GSM8K 0.32, versus 0.88 for PiCO.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Optimize randomly initialized reviewer confidence weights for consistency between capability weights and response scores | PiCO · Chatbot Arena | Spearman's rank correlation: 0.80 -> 0.89 | Helped |
| Assign larger weights to stronger models instead of uniform weights (ablation) | Peer-review ranking · MT-Bench, Chatbot Arena, and AlpacaEval | Spearman's rank correlation: MT-Bench: 0.74 -> 0.75; Chatbot Arena: 0.80 -> 0.82; AlpacaEval: 0.77 -> 0.79 | Helped |
| Assign weights backward by human ranking instead of uniform weights (ablation) | Peer-review ranking · MT-Bench, Chatbot Arena, and AlpacaEval | Spearman's rank correlation: MT-Bench: 0.74 -> 0.70; Chatbot Arena: 0.80 -> 0.72; AlpacaEval: 0.77 -> 0.69 | Hurt |
| Eliminate weaker reviewers, up to 60% of models (ablation) | PiCO · Chatbot Arena, MT-Bench, and AlpacaEval | average training loss: null | Helped |
| Eliminate more than 9 reviewers (ablation) | PiCO · Chatbot Arena, MT-Bench, and AlpacaEval | evaluation performance: null | Hurt |
| Use PiCO ranking instead of existing benchmark ranking | LLM ranking against human preferences · MMLU and GSM8K comparison | Spearman's rank correlation: MMLU: 0.53 -> 0.88; GSM8K: 0.32 -> 0.88 | Helped |
Surprising findings
- Claude-3 (API) is outperformed by PiCO on several metrics but is close on Spearman for Chatbot Arena (0.90 vs 0.90 at ratio 1), showing a strong closed-source reference is nearly matched.
- Removing weaker reviewers improves performance, with the loss minimized when 60% of reviewers are eliminated, and most individual reviewer LLMs yield noisy evaluations.
Limitations
- Evaluated only on three crowdsourced datasets with a pool of 15 mostly older open-source models, and the ground truth is the existing human ranking.
- Consistency optimization uses a straightforward Pearson-based implementation; the authors note more advanced strategies may help.
- Reported results have high variance on some settings (e.g., permutation entropy std up to 0.24) and the method is only validated on text.
Experimental setup
Datasets: MT-Bench Chatbot Arena Conversations AlpacaEval
Benchmarks: MT-Bench Chatbot Arena AlpacaEval MMLU GSM8K
Metrics: Spearman's rank correlation Kendall's tau Permutation entropy Precision@K RBP@K Count inversions Longest increasing subsequence
Baselines: Majority Voting Rating Voting GPTScore (flan-t5-xxl) GPTScore (davinci-002) PandaLM PRD PRE Claude-3 (API)
Builds on: PRD PRE GPTScore PandaLM
Closest papers by meaning
- Chatbot Arena: An Open Platform for Evaluating LLMs by Human Preference (ICML 2024 · similarity 0.90)
- Benchmarking LLMs' Judgments with No Gold Standard (ICLR 2025 · similarity 0.90)
- RocketEval: Efficient automated LLM evaluation via grading checklist (ICLR 2025 · similarity 0.90)
- tinyBenchmarks: evaluating LLMs with fewer examples (ICML 2024 · similarity 0.89)
- LLMEval: A Preliminary Study on How to Evaluate Large Language Models (AAAI 2024 · similarity 0.89)
- Accelerating Unbiased LLM Evaluation via Synthetic Feedback (ICML 2025 · similarity 0.89)
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.