mlsanityData to 7 Oct 2026

Length-Controlled AlpacaEval: A Simple Way to Debias Automatic Evaluators

arXiv 2404.04475 ↗First posted 6 Apr 2024cs.LG cs.AI cs.CLcited by 865

Not on the map. Why?

Cited by 329 in LLM alignment / RLHF / preference optimisation, but placed by its own abstract, which didn’t fit.

Abstract (arXiv)

LLM-based auto-annotators have become a key component of the LLM development process due to their cost-effectiveness and scalability compared to human-based evaluation. However, these auto-annotators can introduce biases that are hard to remove. Even simple, known confounders such as preference for longer outputs remain in existing automated evaluation metrics. We propose a simple regression analysis approach for controlling biases in auto-evaluations. As a real case study, we focus on reducing the length bias of AlpacaEval, a fast and affordable benchmark for instruction-tuned LLMs that uses LLMs to estimate response quality. Despite being highly correlated with human preferences, AlpacaEval is known to favor models that generate longer outputs. We introduce a length-controlled AlpacaEval that aims to answer the counterfactual question: "What would the preference be if the model's and baseline's output had the same length?" To achieve this, we first fit a generalized linear model to predict the biased auto-annotator's preferences based on the mediators we want to control for (length difference) and other relevant features. We then obtain length-controlled preferences by predicting preferences while conditioning the GLM with a zero difference in lengths. Length-controlling not only improves the robustness of the metric to manipulations in model verbosity, but we also find that it increases the Spearman correlation with LMSYS Chatbot Arena from 0.94 to 0.98.

Key facts. Cited by 865 papers in the map (data through 7 Oct 2026). Builds on 16 of its 27 references that are in the map. First posted on arXiv 6 Apr 2024.

Builds on

16 of 27 references.

References · 16

  1. 2022Training language models to follow instructions with human feedbackcited by 15,712
  2. 2023Judging LLM-as-a-Judge with MT-Bench and Chatbot Arenacited by 8,393
  3. 2019BERTScore: Evaluating Text Generation with BERTcited by 5,462
  4. 2023Large Language Models are not Fair Evaluatorscited by 1,050
  5. 2019Evaluating the Factual Consistency of Abstractive Text Summarizationcited by 578
  6. 2023A Long Way to Go: Investigating Length Correlations in RLHFcited by 285
  7. 2017Why We Need New Evaluation Metrics for NLGcited by 232
  8. 2024Disentangling Length from Quality in Direct Preference Optimizationcited by 194
  9. 2023Benchmarking Cognitive Biases in Large Language Models as Evaluatorscited by 147
  10. 2024ODIN: Disentangled Reward Mitigates Hacking in RLHFcited by 133
  11. 2020Evaluating Factuality in Generation with Dependency-level Entailmentcited by 122
  12. 2021A Comprehensive Assessment of Dialog Evaluation Metricscited by 90
4 more
  1. 2023Loose lips sink ships: Mitigating Length Bias in Reinforcement Learning from Human Feedbackcited by 78
  2. 2023Style Over Substance: Evaluation Biases for Large Language Modelscited by 68
  3. 2020Learning an Unreferenced Metric for Online Dialogue Evaluationcited by 52
  4. 2022Spurious Correlations in Reference-Free Evaluation of Text Generationcited by 26

Built on by

Cited by year (2026 to 7 Oct)

Citing papers per year: 2024: 215; 2025: 339; 2026: 311.

865 papers.

Sort:

By area:

  1. 2025DeepSeek-R1: Incentivizing Reasoning Capability in LLMs via Reinforcement Learningcited by 6,146
  2. 2024DeepSeek-V2: A Strong, Economical, and Efficient Mixture-of-Experts Language Modelcited by 1,093
  3. 2024Tulu 3: Pushing Frontiers in Open Language Model Post-Trainingcited by 988
  4. 2024SimPO: Simple Preference Optimization with a Reference-Free Rewardcited by 951
  5. 2024DeepSeek-V3 Technical Reportcited by 724
  6. 2024From Crowdsourced Data to High-Quality Benchmarks: Arena-Hard and BenchBuilder Pipelinecited by 484
  7. 2024From Generation to Judgment: Opportunities and Challenges of LLM-as-a-judgecited by 480
  8. 2024LLMs-as-Judges: A Comprehensive Survey on LLM-based Evaluation Methodscited by 451
  9. 2024Mixture-of-Agents Enhances Large Language Model Capabilitiescited by 420
  10. 2024Magpie: Alignment Data Synthesis from Scratch by Prompting Aligned LLMs with Nothingcited by 341
  11. 2024DeepSeek-Coder-V2: Breaking the Barrier of Closed-Source Models in Code Intelligencecited by 325
  12. 2024OLMoE: Open Mixture-of-Experts Language Modelscited by 291
38 more
  1. 2024LiveBench: A Challenging, Contamination-Limited LLM Benchmarkcited by 242
  2. 2025REINFORCE++: Stabilizing Critic-Free Policy Optimization with Global Advantage Normalizationcited by 233
  3. 2024HelpSteer2: Open-source dataset for training top-performing reward modelscited by 232
  4. 2024Self-Play Preference Optimization for Language Model Alignmentcited by 228
  5. 2025A Survey on Test-Time Scaling in Large Language Models: What, How, Where, and How Well?cited by 215
  6. 2024WildBench: Benchmarking LLMs with Challenging Tasks from Real Users in the Wildcited by 211
  7. 2024Meta-Rewarding Language Models: Self-Improving Alignment with LLM-as-a-Meta-Judgecited by 202
  8. 2025Learning to Reason without External Rewardscited by 199
  9. 2024RM-Bench: Benchmarking Reward Models of Language Models with Subtlety and Stylecited by 168
  10. 2026Entropy-Aware On-Policy Distillation of Language Modelscited by 151
  11. 2024ProSA: Assessing and Understanding the Prompt Sensitivity of LLMscited by 145
  12. 2024Nemotron-4 340B Technical Reportcited by 143
  13. 2024HelpSteer2-Preference: Complementing Ratings with Preferencescited by 129
  14. 2024AudioBench: A Universal Benchmark for Audio Large Language Modelscited by 124
  15. 2025Preference Leakage: A Contamination Problem in LLM-as-a-judgecited by 119
  16. 2024ChartMimic: Evaluating LMM's Cross-Modal Reasoning Capability via Chart-to-Code Generationcited by 118
  17. 2025Checklists Are Better Than Reward Models For Aligning Language Modelscited by 115
  18. 2024Fast Best-of-N Decoding via Speculative Rejectioncited by 112
  19. 2025A Comprehensive Survey in LLM(-Agent) Full Stack Safety: Data, Training and Deploymentcited by 110
  20. 2024Provably Mitigating Overoptimization in RLHF: Your SFT Loss is Implicitly an Adversarial Regularizercited by 109
  21. 2024Unpacking DPO and PPO: Disentangling Best Practices for Learning from Preference Feedbackcited by 107
  22. 2024Exploratory Preference Optimization: Harnessing Implicit Q*-Approximation for Sample-Efficient RLHFcited by 106
  23. 2024Trust or Escalate: LLM Judges with Provable Guarantees for Human Agreementcited by 99
  24. 2024The BiGGen Bench: A Principled Benchmark for Fine-grained Evaluation of Language Models with Language Modelscited by 89
  25. 2024MiLoRA: Harnessing Minor Singular Components for Parameter-Efficient LLM Finetuningcited by 89
  26. 2024Hunyuan-Large: An Open-Source MoE Model with 52 Billion Activated Parameters by Tencentcited by 89
  27. 2024VL-RewardBench: A Challenging Benchmark for Vision-Language Generative Reward Modelscited by 85
  28. 2024Aligning to Thousands of Preferences via System Message Generalizationcited by 84
  29. 2025Reasoning with Sampling: Your Base Model is Smarter Than You Thinkcited by 84
  30. 2024Efficient LLM Scheduling by Learning to Rankcited by 83
  31. 2025Jointly Reinforcing Diversity and Quality in Language Model Generationscited by 79
  32. 2024WildVision: Evaluating Vision-Language Models in the Wild with Human Preferencescited by 76
  33. 2025JudgeLRM: Large Reasoning Models as a Judgecited by 76
  34. 2024RRM: Robust Reward Model Training Mitigates Reward Hackingcited by 75
  35. 2024Value-Incentivized Preference Optimization: A Unified Approach to Online and Offline RLHFcited by 73
  36. 2024MixEval: Deriving Wisdom of the Crowd from LLM Benchmark Mixturescited by 73
  37. 2025InternLM-XComposer2.5-Reward: A Simple Yet Effective Multi-Modal Reward Modelcited by 73
  38. 2024Foundational Autoraters: Taming Large Language Models for Better Automatic Evaluationcited by 72

50 most cited of 865 All 865 (JSON)