Research map / Large language models / Language, safety and interpretability

Heterogeneous Judge-Aware Ranking with Sensitivity, Disagreement, and Confidence

NeurIPS 2026 · Large language models › Language, safety and interpretability · approach: Statistical frameworks for LLM evaluation · 0 citations (OpenAlex)

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 Heterogeneous Judge-Aware (HJA) ranking, an identifiable low-rank-plus-consensus BTL model for multi-judge pairwise comparisons, fit by anchored proximal alternating MLE with fixed-panel Wald inference.

Problem

Pooled ranking models for LLM evaluation treat disagreement among multiple judges as noise, and existing judge-aware models capture only scalar judge sensitivity, so they cannot separate consensus, sensitivity, and structured disagreement or give calibrated uncertainty for ranking contrasts.

Approach

Decompose the judge-specific BTL score matrix as S = gamma mu^T + UV^T with normalization constraints that make consensus mu, judge sensitivity gamma, and orthogonal residual disagreement UV^T identifiable; estimate by constrained MLE with an anchored proximal alternating algorithm and derive asymptotic normality for fixed-panel repeated comparisons.

Key results

• On MT-Bench, HJA reaches 1.00 robustness across noise levels and 0.76 hold-out accuracy vs 0.70 for BTL and 0.70 for JA-Ranking.
• On Chatbot Arena, HJA achieves 0.65 hold-out accuracy vs 0.58 for JA-Ranking and BTL, and 0.82 high-noise robustness vs 0.46 for JA-Ranking.
• On the in-house dataset, JA-Ranking matches or beats HJA under strongest perturbation (0.36 vs 0.33 high-noise robustness for HJA is lower), so HJA does not dominate everywhere.

Surprising findings

• On the in-house dataset HJA does not dominate JA-Ranking under the strongest noisy-judge perturbation (0.33 vs 0.36 for JA-Ranking).

Limitations

• Nonconvex optimization only locally initialized; convergence guarantees are local.
• Severe coordinated judge bias may corrupt the consensus estimate; not designed for adversarial panels.
• Inference assumes fixed comparison graphs and connectivity per judge; random or growing graphs left to future work.

Experimental setup

Datasets: Chatbot Arena MT-Bench UltraFeedback in-house judgment dataset

Benchmarks: Chatbot Arena MT-Bench UltraFeedback in-house judgment dataset

Metrics: MSE Spearman NDCG@N 95% coverage sign accuracy noisy-judge robustness hold-out accuracy near-tie accuracy

Baselines: JA-Ranking pooled BTL unstructured judge-wise BTL + SVD

Builds on: Bradley-Terry-Luce model JA-Ranking low-rank BTL / collaborative ranking

Closest papers by meaning

• From Uncertain Judgments to Calibrated Rankings: Conformal Elo Estimation for LLM Evaluation (NeurIPS 2026 · similarity 0.88)
• JudgeBench: A Benchmark for Evaluating LLM-Based Judges (ICLR 2025 · similarity 0.87)
• Distribution-Calibrated Inference Time Compute for Thinking LLM-as-a-Judge (ICML 2026 · similarity 0.86)
• Multi-Agent Debate for LLM Judges with Adaptive Stability Detection (NeurIPS 2025 · similarity 0.86)
• Aggregating Quantitative Relative Judgments: From Social Choice to Ranking Prediction (NeurIPS 2024 · similarity 0.86)
• Style Outweighs Substance: Failure Modes of LLM Judges in Alignment Benchmarking (ICLR 2025 · similarity 0.86)