AI Researchers Accuse LM Arena of Biased Benchmark Testing Favoring Top Tech Companies

A recent paper co-authored by researchers from AI lab Cohere, Stanford, MIT, and the Allen Institute for AI (Ai2) alleges that LM Arena, the organization behind the popular crowdsourced AI benchmark Chatbot Arena, has been giving preferential treatment to top AI companies. According to the study, LM Arena allowed a select group of companies, including Meta, OpenAI, Google, and Amazon, to privately test multiple variants of their AI models on the platform, but did not disclose the scores of the lowest-performing models. This practice, the authors argue, skewed the leaderboard in favor of these companies, creating an unfair competitive environment. Chatbot Arena, launched in 2023 as an academic research project from UC Berkeley, operates by pit-fighting two AI models in a "battle," where users choose the best response. The results of these battles contribute to the models' scores and ranking on the leaderboard. Many commercial AI companies, including those accused of receiving preferential treatment, have used Chatbot Arena to evaluate their models. However, the paper's findings suggest that this ostensibly impartial benchmark has significant biases. Sara Hooker, Cohere’s VP of AI Research and a co-author of the study, stated in an interview with TechCrunch that only a handful of companies were notified about the availability of private testing, and some received far more opportunities than others. For instance, Meta reportedly tested 27 different model variants between January and March before launching its Llama 4, while only revealing the score of the highest-performing variant. Similarly, the study alleges that OpenAI, Google, and Amazon engaged in similar practices, though to varying degrees. The researchers conducted their investigation over a five-month period, analyzing more than 2.8 million Chatbot Arena battles. They found that certain AI companies had their models appear in disproportionately more battles, which they argue provided these companies with an unfair advantage in data collection and performance optimization. Specifically, the study claims that these companies could boost their models' performance on a related benchmark, Arena Hard, by up to 112% using the additional data. Ion Stoica, LM Arena's co-founder and a professor at UC Berkeley, strongly refuted the study's claims, calling them "full of inaccuracies" and "questionable analysis." In a statement, LM Arena argued that inviting model providers to submit more tests for their models does not equate to unfair treatment, and that it encourages all participants to improve their models through community-driven evaluations. Armand Joulin, a principal researcher at Google DeepMind, also disputed the study's findings, asserting that Google only sent one AI model variant for pre-release testing. Despite these rebuttals, the study highlights the need for greater transparency in AI benchmarking. The authors recommend that LM Arena implement a clear and transparent limit on the number of private tests allowed and disclose the scores from these tests. They also suggest adjusting the sampling rate to ensure that all models appear in the same number of battles. Hooker emphasized that the exact mechanisms allowing certain companies priority access remain murky, and that LM Arena must take responsibility for increasing transparency. The study's reliance on "self-identification" to determine model origins adds a layer of complexity, as models may inaccurately report their company of origin. However, Hooker noted that LM Arena did not initially dispute the preliminary findings shared by the authors. The controversy around Meta's benchmarking practices adds weight to the study's claims. Meta recently faced scrutiny for optimizing one of its Llama 4 models specifically for "conversationality," which led to a high score on Chatbot Arena’s leaderboard. However, the company did not release this optimized model, causing the vanilla version to perform poorly. LM Arena itself acknowledged at the time that Meta should have been more transparent in its approach. Adding to the scrutiny is LM Arena's recent decision to transition from an academic project to a for-profit company, aiming to raise capital from investors. This shift raises questions about the organization's ability to remain unbiased, especially given the potential for corporate influence to affect benchmarking processes. Industry insiders have expressed concerns over the implications of the study. The findings highlight a critical issue in the AI community: the need for fair and transparent benchmarks to accurately assess and compare model performance. Without such standards, smaller and non-major labs may struggle to compete, potentially stifling innovation and diversity in the field. LM Arena, once seen as a neutral platform, now faces pressure to address these issues and rebuild trust with the AI community. Cohere, one of the leading AI research labs, is known for its commitment to ethical and transparent AI practices. Founded in 2019, Cohere has gained recognition for its focus on responsible AI development and collaboration within the academic and industrial sectors. The allegations against LM Arena underscore Cohere's ongoing efforts to promote fairness and integrity in AI benchmarking.