Research map / Large language models / Language, safety and interpretability
Correct Looks Better: Pairwise Comparisons Reveal Accuracy Rankings
Mina Remeli, Moritz Hardt
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 |
|---|---|---|---|
| rank by Bradley-Terry over pairwise judgments instead of direct correctness judging | gpt-oss-20b as weak judge · SimpleQA | Kendall distance to accuracy ranking: 38.2% -> 20% | Improved |
| rank by Bradley-Terry instead of direct judge, weak judge | gpt-oss-20b · SimpleQA | Kendall tau: 0.178 -> 0.667 | Improved |
| rank by Bradley-Terry instead of direct judge, strong judge | o3 · SimpleQA | Kendall tau: 0.922 -> 0.957 | No clear effect |
| rank by Bradley-Terry instead of direct judge, averaged | judge o3 · 5 benchmarks average | Kendall tau: 0.85 -> 0.81 | No clear effect |
| rank by Bradley-Terry instead of direct judge, averaged | judge gpt-oss-20b · 5 benchmarks average | Kendall tau: 0.76 -> 0.80 | Improved |
| add bias correction (style, self-preference) (ablation) | Bradley-Terry ranking · 5 benchmarks | Kendall tau | No clear effect |
| restrict to non-discriminative pairs only (ablation) | Bradley-Terry ranking · 5 benchmarks average | Kendall tau: 0.8 -> 0.68 | Hurt |
| restrict to discriminative pairs only (ablation) | Bradley-Terry ranking · GSM8K | Kendall distance: 7.4% -> 3.4% | Improved |
| restrict to discriminative pairs only (ablation) | Bradley-Terry ranking · BBH | Kendall distance: 7.7% -> 3.8% | Improved |
| append echo (repeated Q-A) to answer A (ablation) | o3 judge · BBH non-discriminative pairs | P(A preferred over B): 0.65 -> 0.29 | Hurt |
| append echo to answer B (ablation) | o3 judge · BBH non-discriminative pairs | P(A preferred over B): 0.65 -> 0.81 | Improved |
| append echo on discriminative pairs (ablation) | o3 judge · BBH discriminative pairs | preference contrast: -0.07 | No clear effect |
Closest papers by meaning
- Bridging Human and LLM Judgments: Understanding and Narrowing the Gap (NeurIPS 2025 · similarity 0.88)
- From Uncertain Judgments to Calibrated Rankings: Conformal Elo Estimation for LLM Evaluation (NeurIPS 2026 · similarity 0.88)
- Prediction-Powered Ranking of Large Language Models (NeurIPS 2024 · similarity 0.88)
- Peering Through Preferences: Unraveling Feedback Acquisition for Aligning Large Language Models (ICLR 2024 · similarity 0.87)
- Heterogeneous Judge-Aware Ranking with Sensitivity, Disagreement, and Confidence (NeurIPS 2026 · similarity 0.87)
- Benchmarking LLMs' Judgments with No Gold Standard (ICLR 2025 · similarity 0.87)
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.
