Research map / Large language models / Language, safety and interpretability
Who can we trust? LLM-as-a-jury for Comparative Assessment
Mengjie Qian, Guangzhi Sun, Mark Gales, Kate Knill
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 judge-specific discriminator to soft BT (BT-σ) | soft BT jury aggregation · SummEval COH | Spearman correlation: 53.94 -> 57.38 | Improved |
| add judge-specific discriminator to soft BT (BT-σ) | soft BT jury aggregation · SummEval ALL | Spearman correlation: 49.40 -> 50.50 | Improved |
| add judge-specific discriminator to soft BT (BT-σ) | soft BT jury aggregation · Topical-Chat CNT | Spearman correlation: 53.87 -> 56.30 | Improved |
| BT-σ instead of supervised temperature scaling (Temp-BT) | jury aggregation · SummEval and Topical-Chat | Spearman correlation: SummEval COH 56.21 -> 57.38 | Mixed |
| use aspect-dependent discriminators (BT-σ-asp) (ablation) | BT-σ · SummEval and Topical-Chat | Spearman correlation: SummEval COH 57.38 -> 57.36 | No clear effect |
| apply discriminators to hard comparisons (hard BT-σ) when inconsistency is high (ablation) | hard BT · Topical-Chat ENG | Spearman correlation: 62.57 -> 67.36 | Improved |
| hard BT-σ instead of soft BT-σ under moderate inconsistency (ablation) | BT-σ · SummEval COH | Spearman correlation: 57.38 -> 53.02 | Hurt |
| use soft instead of hard BT for individual inconsistent judges | single LLM judge · SummEval COH (Qwen2.5-7B-Instruct) | Spearman correlation: 51.42 -> 44.79 | Hurt |
| use soft BT instead of hard BT after aggregating multiple judges | jury aggregation · SummEval ALL | Spearman correlation: 47.34 -> 49.40 | Improved |
| fit a Bradley-Terry model instead of averaging win probabilities | Avg-Prob · SummEval ALL | Spearman correlation: 45.15 -> 49.40 | Improved |
| soft BT instead of hard BT on Topical-Chat ENG | jury aggregation · Topical-Chat ENG | Spearman correlation: 62.57 -> 61.87 | Hurt |
Closest papers by meaning
- JudgeBench: A Benchmark for Evaluating LLM-Based Judges (ICLR 2025 · similarity 0.92)
- Neither Valid nor Reliable? Investigating the Use of LLMs as Judges (NeurIPS 2025 · similarity 0.92)
- Evaluating Judges as Evaluators: The JETTS Benchmark of LLM-as-Judges as Test-Time Scaling Evaluators (ICML 2025 · similarity 0.91)
- JudgeLM: Fine-tuned Large Language Models are Scalable Judges (ICLR 2025 · similarity 0.91)
- Tuning LLM Judge Design Decisions for 1/1000 of the Cost (ICML 2025 · similarity 0.90)
- From Uncertain Judgments to Calibrated Rankings: Conformal Elo Estimation for LLM Evaluation (NeurIPS 2026 · similarity 0.90)
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.
