Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

The Leaderboard Illusion

Shivalika Singh, Yiyang Nan, Alex Wang, Daniel D'Souza, Sayash Kapoor, Ahmet Üstün, Sanmi Koyejo, Yuntian Deng, Shayne Longpre, Noah A. Smith, Beyza Ermis, Marzieh Fadaee, Sara Hooker

TL;DR

The paper analyzes how Chatbot Arena's live leaderboard can be distorted by undisclosed private testing, selective score retraction, and data-access asymmetries that favor proprietary providers. It uses large-scale battle data, simulations of best-of-N submissions, and controlled experiments to show how these practices inflate rankings and enable overfitting to arena dynamics. The authors propose concrete reforms—such as banning score retractions, capping private variants, clarifying deprecation rules, improving sampling fairness, and increasing transparency—to restore scientific integrity and fairness in dynamic benchmarks. Despite acknowledging Arena's community benefits, the study argues that implementing these reforms is essential to ensure leaderboard results reflect genuine, generalizable progress rather than gaming incentives.

Abstract

Measuring progress is fundamental to the advancement of any scientific field. As benchmarks play an increasingly central role, they also grow more susceptible to distortion. Chatbot Arena has emerged as the go-to leaderboard for ranking the most capable AI systems. Yet, in this work we identify systematic issues that have resulted in a distorted playing field. We find that undisclosed private testing practices benefit a handful of providers who are able to test multiple variants before public release and retract scores if desired. We establish that the ability of these providers to choose the best score leads to biased Arena scores due to selective disclosure of performance results. At an extreme, we identify 27 private LLM variants tested by Meta in the lead-up to the Llama-4 release. We also establish that proprietary closed models are sampled at higher rates (number of battles) and have fewer models removed from the arena than open-weight and open-source alternatives. Both these policies lead to large data access asymmetries over time. Providers like Google and OpenAI have received an estimated 19.2% and 20.4% of all data on the arena, respectively. In contrast, a combined 83 open-weight models have only received an estimated 29.7% of the total data. We show that access to Chatbot Arena data yields substantial benefits; even limited additional data can result in relative performance gains of up to 112% on the arena distribution, based on our conservative estimates. Together, these dynamics result in overfitting to Arena-specific dynamics rather than general model quality. The Arena builds on the substantial efforts of both the organizers and an open community that maintains this valuable evaluation platform. We offer actionable recommendations to reform the Chatbot Arena's evaluation framework and promote fairer, more transparent benchmarking for the field

The Leaderboard Illusion

TL;DR

The paper analyzes how Chatbot Arena's live leaderboard can be distorted by undisclosed private testing, selective score retraction, and data-access asymmetries that favor proprietary providers. It uses large-scale battle data, simulations of best-of-N submissions, and controlled experiments to show how these practices inflate rankings and enable overfitting to arena dynamics. The authors propose concrete reforms—such as banning score retractions, capping private variants, clarifying deprecation rules, improving sampling fairness, and increasing transparency—to restore scientific integrity and fairness in dynamic benchmarks. Despite acknowledging Arena's community benefits, the study argues that implementing these reforms is essential to ensure leaderboard results reflect genuine, generalizable progress rather than gaming incentives.

Abstract

Measuring progress is fundamental to the advancement of any scientific field. As benchmarks play an increasingly central role, they also grow more susceptible to distortion. Chatbot Arena has emerged as the go-to leaderboard for ranking the most capable AI systems. Yet, in this work we identify systematic issues that have resulted in a distorted playing field. We find that undisclosed private testing practices benefit a handful of providers who are able to test multiple variants before public release and retract scores if desired. We establish that the ability of these providers to choose the best score leads to biased Arena scores due to selective disclosure of performance results. At an extreme, we identify 27 private LLM variants tested by Meta in the lead-up to the Llama-4 release. We also establish that proprietary closed models are sampled at higher rates (number of battles) and have fewer models removed from the arena than open-weight and open-source alternatives. Both these policies lead to large data access asymmetries over time. Providers like Google and OpenAI have received an estimated 19.2% and 20.4% of all data on the arena, respectively. In contrast, a combined 83 open-weight models have only received an estimated 29.7% of the total data. We show that access to Chatbot Arena data yields substantial benefits; even limited additional data can result in relative performance gains of up to 112% on the arena distribution, based on our conservative estimates. Together, these dynamics result in overfitting to Arena-specific dynamics rather than general model quality. The Arena builds on the substantial efforts of both the organizers and an open community that maintains this valuable evaluation platform. We offer actionable recommendations to reform the Chatbot Arena's evaluation framework and promote fairer, more transparent benchmarking for the field
Paper Structure (44 sections, 1 theorem, 29 equations, 19 figures, 4 tables)

This paper contains 44 sections, 1 theorem, 29 equations, 19 figures, 4 tables.

Table of Contents

  1. Introduction
  2. Overview of Methodology
  3. Introduction to the Arena Score
  4. Results: Impact of Private Testing and Selective Retraction on Arena Scores
  5. Preferred Providers Frequently Use Private Testing
  6. Simulated Experiments on Private Testing and Retraction
  7. Real-world Chatbot Arena Experiment
  8. Results: Impact of Data Access Asymmetries on Arena Scores
  9. Disparity in access to Chatbot Arena Data
  10. Risk of Potential Overfitting
  11. Results: Impact of Model Deprecation on Arena Scores
  12. Transitivity Under Changing Evaluation Conditions
  13. Sparse Battle History Risks
  14. Recommendations and Guidelines for Improving Leaderboards
  15. Limitations
  16. ...and 29 more sections

Key Result

Theorem 1

For every $N \geq 2$,

Figures (19)

  • Figure 1: Overview of key insights. We investigate the prevalence of undisclosed private testing and selective score reporting on the Arena (\ref{['sec:impact-of-ability-to-pricately-test']}), and highlight significant data access disparities between proprietary and open-source providers (\ref{['sec-disparity-in-data-access']}). These disparities enable overfitting to the Arena (\ref{['sec:risk_of_potential_overfitting']}). Furthermore, model deprecation practices lack transparency, with many models silently deprecated without any notification to providers. We demonstrate how these deprecations contribute to unreliable rankings on the leaderboard (\ref{['sec:battle-connecitivty']}).
  • Figure 2: Number of public models vs. maximum arena score per provider. Marker size indicates total number of battles played. Proprietary model providers tend to achieve higher leaderboard scores, which appear to correlate with both the number of models they release and the number of Arena battles played. While model capability is an important factor, we explore in \ref{['sec:impact-of-ability-to-pricately-test']} and \ref{['sec-measuring-overfititng']} how increased exposure to the Arena (through more models and battles) may confer additional advantages, such as better model selection or adaptation to the evaluation distribution. This figure summarizes publicly disclosed results as of April 23rd, 2025.
  • Figure 3: Volume of Arena battles involving proprietary, open-weight, and fully open-source model providers from January 2024 to March 2025, based on leaderboard-stats. Proprietary models consistently received the largest share of data---ranging from 54.3% to 70.1%. Open-weight and fully open-source models receive significantly less data, in some cases receiving less than half the amount of data as proprietary developers. This imbalance in data access exacerbates the performance gap, reinforcing unequal access to high-quality in-distribution data.
  • Figure 4: Data availability to model providers. We observe large differences in data access between providers, with 61.4% of all data going to proprietary providers.
  • Figure 5: Maximum observed sampling rate for models from different providers. The sampling rate determines the amount of times a model is shown to everyday users, and the amount of data a provider receives. We observe large discrepancies across providers, with substantially higher sampling rates for OpenAI, Google, xAI, and Meta compared to others.
  • ...and 14 more figures

Theorems & Definitions (3)

  • Theorem 1
  • Proof 1
  • Remark 1