mlsanityData to 7 Oct 2026

Margin-Adaptive Confidence Ranking for Reliable LLM Judgement

arXiv 2605.15416 ↗First posted 14 May 2026cs.LG cs.AIcited by 1

Not on the map. Why?

Abstract (arXiv)

Jung et al. (2025) introduce a hypothesis testing framework for guaranteeing agreement between large language models (LLMs) and human judgments, relying on the assumption that the model's estimated confidence is monotonic with respect to human-disagreement risk. In practice, however, this assumption may be violated, and the generalization behavior of the confidence estimator is not explicitly analyzed. We mitigate these issues by learning a dedicated confidence estimator instead of relying on heuristic confidence signals. Our approach leverages simulated annotator diversity and a margin-based ranking formulation to explicitly model how confidently an LLM distinguishes between human-agreement and human-disagreement cases. We further derive generalization guarantees for this estimator, revealing a margin-dependent trade-off that informs the design of an adaptive estimator training procedure. When integrated into fixed-sequence testing, the learned confidence estimator yields improved ranking accuracy and empirically strengthens the monotonic relationship between confidence and disagreement risk, leading to higher success rates in satisfying target agreement levels across multiple datasets and judge models.

Key facts. Cited by 1 paper in the map (data through 7 Oct 2026). Builds on 51 of its 63 references that are in the map. First posted on arXiv 14 May 2026.

Builds on

51 of 63 references.

References · 51

  1. 2023Judging LLM-as-a-Judge with MT-Bench and Chatbot Arenacited by 8,393
  2. 2022Training a Helpful and Harmless Assistant with Reinforcement Learning from Human Feedbackcited by 3,634
  3. 2020Learning to summarize from human feedbackcited by 2,562
  4. 2023G-Eval: NLG Evaluation using GPT-4 with Better Human Alignmentcited by 2,205
  5. 2023FActScore: Fine-grained Atomic Evaluation of Factual Precision in Long Form Text Generationcited by 1,259
  6. 2023Large Language Models are not Fair Evaluatorscited by 1,050
  7. 2023Can LLMs Express Their Uncertainty? An Empirical Evaluation of Confidence Elicitation in LLMscited by 946
  8. 2023AlpacaFarm: A Simulation Framework for Methods that Learn from Human Feedbackcited by 738
  9. 2023Can Large Language Models Be an Alternative to Human Evaluations?cited by 729
  10. 2023ChatEval: Towards Better LLM-based Evaluators through Multi-Agent Debatecited by 724
  11. 2024LLM Evaluators Recognize and Favor Their Own Generationscited by 651
  12. 2024Prometheus 2: An Open Source Language Model Specialized in Evaluating Other Language Modelscited by 431
38 more
  1. 2023GPTScore: Evaluate as You Desirecited by 357
  2. 2023JudgeLM: Fine-tuned Large Language Models are Scalable Judgescited by 351
  3. 2023PandaLM: An Automatic Evaluation Benchmark for LLM Instruction Tuning Optimizationcited by 311
  4. 2017A PAC-Bayesian Approach to Spectrally-Normalized Margin Bounds for Neural Networkscited by 309
  5. 2024Replacing Judges with Juries: Evaluating LLM Generations with a Panel of Diverse Modelscited by 301
  6. 2022Conformal Risk Controlcited by 290
  7. 2023Evaluating Large Language Models at Evaluating Instruction Followingcited by 290
  8. 2021All That's 'Human' Is Not Gold: Evaluating Human Evaluation of Generated Textcited by 287
  9. 2023AlignScore: Evaluating Factual Consistency with a Unified Alignment Functioncited by 284
  10. 2021Distribution-Free, Risk-Controlling Prediction Setscited by 278
  11. 2023Fine-tuning Language Models for Factualitycited by 219
  12. 2023Benchmarking Foundation Models with Language-Model-as-an-Examinercited by 172
  13. 2023Benchmarking Cognitive Biases in Large Language Models as Evaluatorscited by 147
  14. 2024Language Models with Conformal Factuality Guaranteescited by 128
  15. 2023Human-like Summarization Evaluation with ChatGPTcited by 126
  16. 2023R-Tuning: Instructing Large Language Models to Say `I Don't Know'cited by 106
  17. 2023LLM-Eval: Unified Multi-Dimensional Automatic Evaluation for Open-Domain Conversations with Large Language Modelscited by 100
  18. 2024Trust or Escalate: LLM Judges with Provable Guarantees for Human Agreementcited by 99
  19. 2024Unfamiliar Finetuning Examples Control How Language Models Hallucinatecited by 85
  20. 2023Style Over Substance: Evaluation Biases for Large Language Modelscited by 68
  21. 2024Mitigating LLM Hallucinations via Conformal Abstentioncited by 56
  22. 2024WPO: Enhancing RLHF with Weighted Preference Optimizationcited by 53
  23. 2023ICE-Score: Instructing Large Language Models to Evaluate Codecited by 50
  24. 2024Conformal Alignment: Knowing When to Trust Foundation Models with Guaranteescited by 50
  25. 2024Multicalibration for Confidence Scoring in LLMscited by 44
  26. 2022Enhancing Adversarial Training with Second-Order Statistics of Weightscited by 42
  27. 2020How does Weight Correlation Affect the Generalisation Ability of Deep Neural Networkscited by 32
  28. 2025Calibrating LLM Confidence by Probing Perturbed Representation Stabilitycited by 23
  29. 2025SteerConf: Steering LLMs for Confidence Elicitationcited by 20
  30. 2025Enhancing Robust Fairness via Confusional Spectral Regularizationcited by 11
  31. 2025Adaptive Prediction-Powered AutoEval with Reliability and Efficiency Guaranteescited by 9
  32. 2024HREF: Human Response-Guided Evaluation of Instruction Following in Language Modelscited by 8
  33. 2025Multi-Agent LLM Judge: automatic personalized LLM judge design for evaluating natural language generation applicationscited by 7
  34. 2022Weight Expansion: A New Perspective on Dropout and Generalizationcited by 6
  35. 2025CLEV: LLM-Based Evaluation Through Lightweight Efficient Voting for Free-Form Question-Answeringcited by 6
  36. 2025Bridging Human and LLM Judgments: Understanding and Narrowing the Gapcited by 6
  37. 2026Towards A Unified PAC-Bayesian Framework for Norm-based Generalization Boundscited by 4
  38. 2025Through the Judge's Eyes: Inferred Thinking Traces Improve Reliability of LLM Raterscited by 2

50 most cited of 51 All 51 (JSON)

Built on by

Cited by year (2026 to 7 Oct)

Citing papers per year: 2026: 1.

1 paper.

  1. 2026Localize-Then-Decide Guarantees for LLM Judgmentscited by 0