PaperlayerView exact revision on arXiv

Source-linked AI summary

Chatbot Arena: An Open Platform for Evaluating LLMs by Human Preference

Wei-Lin Chiang, Lianmin Zheng, Ying Sheng, Anastasios Nikolas Angelopoulos, Tianle Li, Dacheng Li, Hao Zhang, Banghua Zhu, Michael Jordan, Joseph E. Gonzalez, Ion Stoica

arXiv:2403.04132v1cs.AIcs.CL

TL;DR

Evaluating whether LLMs align with human preferences remains difficult because existing benchmarks inadequately represent diverse, open-ended use. This paper introduces Chatbot Arena, which crowdsources anonymous pairwise comparisons, applies statistical ranking and efficient sampling, and analyzes the resulting data. The platform accumulated over 240K votes, with prompts that are diverse and discriminating and crowd judgments highly consistent with expert evaluations.

  • Problem

    Existing benchmarks often fail to capture nuanced, diverse, open-ended tasks and alignment with human preferences, while live evaluation also requires scalable ranking and reliable human judgments.

  • Method

    Chatbot Arena uses anonymous randomized battles, unrestricted user prompts, pairwise preference votes, statistical ranking models, and efficient model-pair sampling.

  • Results

    The platform received over 240K votes from about 90K users across over 100 languages, and analyses found diverse, discriminating prompts with crowd votes highly consistent with expert evaluations.

  • Takeaways & Limitations

    Chatbot Arena provides an open platform and dataset for evaluating LLMs through crowdsourced human preferences, including a planned release of over 100K pairwise votes.

  • Takeaways & Limitations

    The user base may be biased toward LLM hobbyists and researchers, online-chat prompts may not represent production or specialized-domain usage, and the study does not evaluate safety.

Abstract

from arXiv · show

Large Language Models (LLMs) have unlocked new capabilities and applications; however, evaluating the alignment with human preferences still poses significant challenges. To address this issue, we introduce Chatbot Arena, an open platform for evaluating LLMs based on human preferences. Our methodology employs a pairwise comparison approach and leverages input from a diverse user base through crowdsourcing. The platform has been operational for several months, amassing over 240K votes. This paper describes the platform, analyzes the data we have collected so far, and explains the tried-and-true statistical methods we are using for efficient and accurate evaluation and ranking of models. We confirm that the crowdsourced questions are sufficiently diverse and discriminating and that the crowdsourced human votes are in good agreement with those of expert raters. These analyses collectively establish a robust foundation for the credibility of Chatbot Arena. Because of its unique value and openness, Chatbot Arena has emerged as one of the most referenced LLM leaderboards, widely cited by leading LLM developers and companies. Our demo is publicly available at \url{https://chat.lmsys.org}.

1. Introduction

Existing benchmarks often miss open-ended, real-world human preferences, motivating Chatbot Arena, a live crowdsourced platform that uses pairwise votes and statistical ranking methods. Analyses report diverse, discriminating prompts and strong crowd–expert agreement, supported by more than 240K votes.

  • Motivation: Static, ground-truth benchmarks inadequately capture open-ended usage, contamination risks, difficult-to-define ground truth, and human preferences.The paper motivates an open, live, human-preference platform for more realistic evaluation.
  • Platform: Chatbot Arena collects anonymous pairwise votes on responses to users’ unrestricted prompts, producing fresh questions that reflect real-world applications.Models are revealed only after voting, and statistical tools estimate rankings efficiently.
  • Findings: Crowdsourced questions are sufficiently diverse and discriminating, while crowd votes are highly consistent with expert evaluations.These findings are presented as evidence supporting Chatbot Arena’s credibility.
  • Scale: More than 240K votes came from about 90K users across over 100 languages as of January 2024.The platform had operated since April 2023 and included over 50 state-of-the-art models.
  • Analysis and release: The platform analyzes prompt diversity, prompt difficulty, vote quality, and human feedback to assess evaluation credibility.The contributions include an in-depth data analysis and a planned public release of over 100K pairwise votes.
  • Ranking: An adaptive model-pair sampling algorithm improves sample efficiency, sometimes substantially, while retaining statistically valid ranking evaluation.The method actively selects which model pairs users compare.

2. Related Work

The related work contrasts prevalent static, ground-truth benchmarks with live human-in-the-loop evaluation, ranking research, and existing human-preference datasets. Chatbot Arena follows the live-evaluation direction at a substantially larger user scale and adds preference data to crowdsourced conversations.

  • LLM Benchmarks: Common LLM benchmarks are predominantly static, ground-truth-based tasks with predefined answers, covering language understanding, mathematics, coding, and reasoning.Examples include MMLU, HellaSwag, GSM-8K, BigBench, and AGIEval.
  • Risks of Static Benchmarks: Static benchmarks face contamination, saturation, overfitting, and limited human alignment, motivating live human-in-the-loop evaluation.DynaBench identifies these challenges and recommends live benchmarks for classical NLP evaluation.
  • Risks of Static Benchmarks: Chatbot Arena adopts the live-benchmark spirit for chatting with LLMs while operating at a significantly larger user scale.The paper distinguishes its focus on conversational LLM evaluation from prior classical NLP settings.
  • Ranking System: The paper contributes techniques for accelerating ranking convergence and detecting abnormal behavior within established ranking-system research.Related approaches include probability models, rank elicitation, online experiment design, and Elo ratings.
  • Human Preference Dataset: Existing human-preference resources include crowdsourced conversational datasets and synthetic preference approximations, whereas LMSYS-Chat-1M lacks preference labels.The paper positions Chatbot Arena as a source of pairwise human-preference data.

3. Human Preference Data Collection

Chatbot Arena uses a low-friction pairwise interface to crowdsource preferences from unrestricted user prompts and anonymous model battles. The resulting data spans many users, models, and languages, with sampling, filtering, and moderation procedures supporting collection quality.

  • Interface Design: Pairwise comparison lets users choose the better of two responses instead of assigning absolute scores, reducing the need for a consistent grading rubric.This design addresses variability across many human evaluators.
  • Battle Procedure: Each battle samples two anonymous models, accepts any user prompt, presents responses side by side, and reveals identities only after voting.Users may continue chatting when the first turn does not establish a preference.
  • Collected Data: As of January 2024, the dataset contained around 240K votes from over 90K users across more than 50 models and over 100 languages.English accounted for 77% of conversations, Chinese for 5%, and each of several other listed languages for less than 2%.
  • Sampling: Non-uniform sampling concentrates votes on similarly performing model pairs, reducing votes needed to reach stable results.The paper later compares this approach with adaptive sampling and random sampling.
  • Data Quality: Identity filtering removes conversations naming models or companies, while moderation flags unsafe requests representing 3% of total requests.These procedures support anonymity and help avoid misuse.

4. From Pairwise Comparisons to Rankings

The paper converts crowdsourced pairwise model comparisons into win matrices, scores, and ranks. It uses Bradley-Terry coefficients to model pairwise outcomes while allowing extensions beyond binary preferences.

  • Pairwise comparisons: Pairwise comparisons are represented as a dataset over model pairs and human responses, from which a ranking over all M models is recovered.The sequential formulation serves a selected model pair and observes the human preference response.
  • Feedback model: The framework primarily uses binary outcomes but can generalize to graded preferences or ties.The response variable is introduced as potentially taking values beyond the binary case.
  • Win matrix: The win matrix records the expected human preference for every ordered model pair.For binary feedback, each entry is the probability that the human prefers the second model in the displayed pair.
  • Scores and ranks: Model scores are defined as an M-dimensional vector whose larger values indicate better models, and ranks count models with higher scores.The best model receives rank 1, with ties sharing that rank.
  • Bradley-Terry scores: Bradley-Terry coefficients provide a standard score function by modeling the probability that one model beats another.The coefficients are identifiable only up to an additive constant, so one coefficient is fixed to zero.

5. Efficient Approximate Ranking

The paper estimates win matrices and Bradley-Terry scores from adaptively sampled comparisons, then quantifies ranking uncertainty and detects anomalous users. Its confidence procedures rely on unbiased estimation, asymptotic intervals, and sequential tests.

  • Win matrix estimation: The win-matrix estimator averages inverse-probability-weighted observations from sampled model pairs.Each weighted observation has expectation equal to the corresponding population win-matrix entry, making the estimator unbiased.
  • Score estimation: Bradley-Terry scores are estimated by reweighted maximum likelihood, with inverse sampling probabilities targeting a uniform distribution over model pairs.The reweighting corrects for the sampling distribution used to collect comparisons.
  • Uncertainty quantification: Confidence intervals for Bradley-Terry coefficients use sandwich robust standard errors, selected over pivot bootstrap intervals because simulations found smaller intervals in large samples.The procedure is used to quantify uncertainty in the estimated coefficients.
  • Approximate rankings: Approximate rankings account for score-estimation uncertainty through confidence sets that can bound whether a model outranks another.The construction uses chi-squared intervals based on the central limit theorem and sandwich variance estimates.
  • Active sampling: Active sampling selects model pairs according to their expected reduction in confidence-interval size.This focuses comparisons where additional data are expected to reduce ranking uncertainty.
  • Detecting anomalous users: Anomalous-user detection compares a new user’s ratings with historical rating distributions and combines sequential p-values with Fisher’s test and a Bonferroni variant.Under the stated exchangeability null, each component p-value is valid and their dependence is asymptotically negligible.

6. Data Analysis

Chatbot Arena analyzes whether crowdsourced prompts and votes can support credible LLM evaluation. The analyses find broad, discriminating prompts and substantial agreement between crowd users, experts, and automated judges.

  • 6.1. Topic Modeling on User Prompts: The prompt-analysis pipeline uses topic modeling to examine whether user questions reflect diverse, real-world LLM use cases.It combines embeddings, dimensionality reduction, and density-based clustering.
  • 6.1. Topic Modeling on User Prompts: 600 clusters cover topics including poetry, coding, math, and medical queries, with the largest cluster representing only 1% of prompts.The remaining clusters quickly fall below 0.5%, indicating a long-tail distribution with low similarity between clusters.
  • 6.2. Can Arena Prompts Distinguish Models?: GPT-4 wins up to 97% on coding and reasoning clusters but below 60% on less problem-solving-oriented clusters against Llama-2-70b-chat.The comparison uses 30 prompts from seven topic clusters and LLM-as-judge evaluation.
  • 6.2. Can Arena Prompts Distinguish Models?: Arena prompts can differentiate model strengths because performance varies substantially across topic clusters.The results show that model comparisons depend on the kinds of tasks represented by the prompts.
  • 6.3. Validating Vote Quality: 72% to 83% agreement between crowd users and experts supports the quality of crowdsourced votes.Expert-versus-expert agreement is 79.4% and 89.8%, while the 5%-10% gap is attributed mainly to crowd mistakes or overlooked factual errors.

7. Experiments

The experiments evaluate interval coverage, adaptive sampling, and anomalous-user detection for Arena’s ranking pipeline. Simulations support the interval behavior, adaptive sampling improves estimation efficiency, and the detection method identifies many anomalous users.

  • 7.1. Ranking system: The interval-coverage experiment simulates Bradley-Terry coefficients and evaluates coverage and average interval width across 20 trials.The simulation varies the number of models and samples, because true coefficients cannot be evaluated directly on real data.
  • 7.1. Ranking system: Coverage centers around 1 −α regardless of the number of models, while intervals become wider as more models are included.These findings are reported for the uncorrected intervals in the simulation.
  • 7.1. Ranking system: Adaptive sampling improves interval-width estimates for both the win matrix and Bradley-Terry coefficients compared with random sampling.The improvement is substantial for win-matrix estimation but more subtle on the score scale.
  • 7.2. Anomalous Users Detection: 90% true-positive and 60-70% true-negative rates demonstrate effective anomalous-user detection.False negatives arise from users whose behavior is not consistently abnormal.

8. Discussion

The authors identify scope limitations in Arena’s user base, prompt source, and evaluation target. They also propose broader future directions, including multimodal and agent-based evaluations and more rigorous harmful-user detection.

  • Limitations: Arena’s users may be biased toward LLM hobbyists and researchers eager to test and evaluate new models.This demographic concentration may produce a biased user distribution.
  • Limitations: Prompts collected through the online chat interface may not represent production environments or specialized domains.The authors therefore note that the prompt distribution may be skewed relative to real-world usage.
  • Limitations: The study evaluates LLM helpfulness but does not assess safety, motivating a parallel mechanism for safety evaluation.The authors explicitly recognize the necessity of such a mechanism.
  • Future Directions: Future work will develop topic leaderboards and dedicated multimodal and agent-based sections for more dynamic, complex tasks.The authors also plan to improve harmful-user detection using nonnegative supermartingales and E-values.

9. Conclusion

The paper presents Chatbot Arena as an open platform for crowdsourced, pairwise human-preference evaluation of LLMs. It analyzes prompt and vote quality, develops ranking methods, and releases a large preference dataset for future research.

  • 9. Conclusion: Chatbot Arena evaluates LLMs through crowdsourced pairwise human preferences on an open platform.The paper presents the platform as a system for collecting and analyzing user prompts and preference votes.
  • 9. Conclusion: The paper validates the diversity and quality of crowdsourced prompts and preference votes while developing efficient model sampling and ranking algorithms.These analyses support the platform’s evaluation framework.
  • 9. Conclusion: A dataset containing 100K pairwise preference votes will be released for future research.

A. Confidence Interval Simulation Study

The simulation study compares bootstrap confidence intervals with the sandwich estimator and finds them largely similar, with good practical coverage and width across tested conditions.

  • Bootstrap and sandwich confidence intervals were often identical to the naked eye.The study nevertheless observed some differences in experiments.
  • 213576 data points were used in a replay study comparing the two interval methods.
  • Both intervals showed good coverage and width across many choices of γ and model strength.The simulations used the same beta-generating process as the main text with γ = 2.

B. The Nonparametric Bradley-Terry Model

The nonparametric Bradley-Terry model ranks models by averaging path-based log odds, while retaining validity beyond the original model’s binary, transitive, and correctly specified setting.

  • The model defines G(m) as all paths leading to model m and averages path-sums of log odds over these paths.Each path is a chain of model pairings that leads to m.
  • The path construction uses directed pairings and their reversals to form chains ending at the target model.The set B includes observed pairings and reversed pairings.
  • The resulting score is a smooth function of θ with an explicitly derived derivative.
  • The original Bradley-Terry model estimates coefficients ξ1, ..., ξM from observed binary outcomes and modeled win probabilities.
  • All path sums equal ξm − ξg1,1, linking the nonparametric score to Bradley-Terry coefficient differences.
  • Under a well-specified parametric Bradley-Terry model, the nonparametric version exactly recovers the Bradley-Terry coefficients.
  • The nonparametric model remains statistically valid when outcomes are nonbinary, win rates are non-transitive, or the logistic model is misspecified.Its coefficient can be computed using equation (16).
  • The p-value construction is valid under exchangeability, satisfying P(pi ≤ t) ≤ t.The proof uses the uniform distribution of the test statistic among the ranks of H′.

D. Sample Prompts

The sample prompts illustrate how coding and reasoning tasks can separate models, while movie-recommendation and travel-planning prompts may not distinguish them.

  • D.1. Prompt Examples From Different Clusters: A snowman-game prompt asks for a Hangman-like game with a dictionary and three difficulty levels.
  • D.1. Prompt Examples From Different Clusters: The GPT-4-Turbo response provides a correct implementation and helpful explanation, whereas Llama-2-70b-chat provides no program and a less helpful response.
  • D.1. Prompt Examples From Different Clusters: The snowman-game implementation includes random word selection, guessed-letter tracking, a ten-guess limit, and snowman drawing output.
  • D.1. Prompt Examples From Different Clusters: The example response also describes optional sound effects, refinement through testing and feedback, and standalone packaging for distribution.
  • D.1. Prompt Examples From Different Clusters: The displayed program output is presented as demonstrating GPT-4-0613’s correctness and greater helpfulness and relevance.
  • D.1. Prompt Examples From Different Clusters: A Christmas-tree coding prompt provides another example in which GPT-4-0613 and Llama-2-70b-chat produce code outputs for comparison.
  • D.1. Prompt Examples From Different Clusters: Movie-recommendation and travel-itinerary prompts are described as nondiscriminating because both assistants provide correct or equally helpful responses.
Loading 2403.04132v1…