Calibrate, Don't Curate: Label-Efficient Estimation from Noisy LLM Judges

• URL: http://arxiv.org/abs/2605.09702v1
• Date: Sun, 10 May 2026 18:49:58 GMT
• Title: Calibrate, Don't Curate: Label-Efficient Estimation from Noisy LLM Judges
• Abstract summary: We show that in multi-judge evaluation with labeled calibration data, do not discard weak judges by accuracy alone.<n>We explain this reversal with oracle analyses showing that the optimal calibrated risk under proper scoring rules cannot increase when additional judge signals are made available.
• Score: 3.974554103907186
• License: http://creativecommons.org/licenses/by/4.0/
• Abstract: Multi-judge evaluation is increasingly used to assess LLMs and reward models, and the prevailing heuristic is to curate: keep the most accurate judges and discard weaker ones. We show that this heuristic can reverse when the target is not point accuracy, but calibrated probabilistic evaluation from a labeled calibration set. Holding the aggregation and calibration procedures fixed, we compare accuracy-ranked top-$k$ judge selection with using the full judge panel. Across four labeled pairwise-evaluation benchmarks spanning LLM-as-judge and reward-model settings, the calibrated full panel consistently outperforms accuracy-based selection. On RewardBench2, retaining all judges achieves negative log-likelihood (NLL) of $0.006$ versus $0.013$ under top-5 selection, halving the calibration error. This advantage persists after judge-family deduplication and against stronger same-pipeline subset search. We explain this reversal with oracle analyses showing that the optimal calibrated risk under proper scoring rules cannot increase when additional judge signals are made available, and that even below-chance judges can be useful when their biases are learnable and their signals are non-redundant. The resulting operating principle is simple: in multi-judge evaluation with labeled calibration data, do not discard weak judges by accuracy alone; keep them when they are parseable, non-redundant, and calibratable.

Related papers

• Can We Trust LLM Judges: A Study of Capability-Dependent Biases and Multi-Judge Ensemble for Bias Calibration [0.5294698352039444]
  We study absolute scoring tasks, which mirror more realistic use cases.<n>We show that a model's task accuracy strongly predicts its judging accuracy.<n>More capable examinee models consistently receive more lenient judgments from all judges.
  arXiv  Detail & Related papers  (2026-09-10T00:25:39Z)
• RecurSE: Bounded Recursive Self-Evaluation for LLM Rubric Judges [48.831446371402414]
  RecurSE pairs a trainable judge with a synchronized policy-copy checker that audits the judge's reasoning against meta-rubrics to supply a scalar process reward.<n>We show that RecurSE achieves consistent generalization gains across held-out medical, pairwise, summarization, and professional benchmarks.
  arXiv  Detail & Related papers  (2026-08-25T08:35:04Z)
• DA-RAC: Distance-Aware Calibration of LLM Judges for Trustworthy AI Auditing [6.639255145889912]
  DA-RAC is a distance-aware reference-anchored calibration method for LLM judges.<n>It improves calibration and reduces false-pass risk relative to zero-shot, chain-of-thought evaluation, and static-anchor baselines.
  arXiv  Detail & Related papers  (2026-08-15T00:29:25Z)
• CyclicJudge: Mitigating Judge Bias Efficiently in LLM-based Evaluation [6.3121191919394475]
  This work introduces a variance decomposition that partitions benchmark score variance into scenario, generation, judge, and residual components.<n>Based on this analysis, CyclicJudge, a round-robin assignment of judges to scenarios, is demonstrated to be the optimal strategy for a fixed judge-call budget.
  arXiv  Detail & Related papers  (2026-03-02T13:46:32Z)
• Who can we trust? LLM-as-a-jury for Comparative Assessment [42.32900791516691]
  Large language models (LLMs) are increasingly applied as automatic evaluators for natural language generation assessment.<n>LLMs judges vary substantially in performance across tasks and aspects, and their judgment probabilities may be biased and inconsistent.<n>We propose BT-sigma, a judge-aware extension of the Bradley-Terry model that introduces a discriminator parameter for each judge to jointly infer item rankings and judge reliability from pairwise comparisons alone.
  arXiv  Detail & Related papers  (2026-02-18T17:04:02Z)
• A Judge-Aware Ranking Framework for Evaluating Large Language Models without Ground Truth [4.9467757325435775]
  evaluating large language models (LLMs) on open-ended tasks is increasingly done via the LLM-as-a-judge paradigm.<n>Treating all judges equally can yield biased leaderboards and misleading uncertainty estimates.<n>We propose a judge-aware ranking framework that extends the Bradley-Terry-Luce model by introducing judge-specific discrimination parameters.
  arXiv  Detail & Related papers  (2026-01-29T15:01:28Z)
• Reference-Free Rating of LLM Responses via Latent Information [53.463883683503106]
  We study the common practice of asking a judge model to assign Likert-scale scores to free-text responses.<n>We then propose and evaluate Latent Judges, which derive scalar ratings from internal model signals.<n>Across a broad suite of pairwise and single-rating benchmarks, latent methods match or surpass standard prompting.
  arXiv  Detail & Related papers  (2025-09-29T12:15:52Z)
• Decision from Suboptimal Classifiers: Excess Risk Pre- and Post-Calibration [52.70324949884702]
  We quantify the excess risk incurred using approximate posterior probabilities in batch binary decision-making.<n>We identify regimes where recalibration alone addresses most of the regret, and regimes where the regret is dominated by the grouping loss.<n>On NLP experiments, we show that these quantities identify when the expected gain of more advanced post-training is worth the operational cost.
  arXiv  Detail & Related papers  (2025-03-23T10:52:36Z)
• JudgeBench: A Benchmark for Evaluating LLM-based Judges [61.048125269475854]
  JudgeBench is a benchmark for evaluating LLM-based judges on challenging response pairs spanning knowledge, reasoning, math, and coding.<n>Our comprehensive evaluation on a collection of prompted judges, fine-tuned judges, multi-agent judges, and reward models shows that JudgeBench poses a significantly greater challenge than previous benchmarks.
  arXiv  Detail & Related papers  (2024-10-16T17:58:19Z)
• Calibration by Distribution Matching: Trainable Kernel Calibration Metrics [56.629245030893685]
  We introduce kernel-based calibration metrics that unify and generalize popular forms of calibration for both classification and regression. These metrics admit differentiable sample estimates, making it easy to incorporate a calibration objective into empirical risk minimization. We provide intuitive mechanisms to tailor calibration metrics to a decision task, and enforce accurate loss estimation and no regret decisions.
  arXiv  Detail & Related papers  (2023-10-31T06:19:40Z)
• Beyond calibration: estimating the grouping loss of modern neural networks [68.8204255655161]
  Proper scoring rule theory shows that given the calibration loss, the missing piece to characterize individual errors is the grouping loss. We show that modern neural network architectures in vision and NLP exhibit grouping loss, notably in distribution shifts settings.
  arXiv  Detail & Related papers  (2022-10-28T07:04:20Z)

This list is automatically generated from the titles and abstracts of the papers in this site.

This site does not guarantee the quality of this site (including all information) and is not responsible for any consequences.