LM Arena is a benchmark platform for large language models that was created in May 2023. It features anonymous randomized battles in a crowdsourced manner, where two LLMs are pitted against each other for the same prompt, and the responses are presented to the user in a blind fashion. LM Arena has become an industry standard, with the CEO of Google highlighting Gemini 2.5 Flash as the top-rated model on the leaderboard during a recent keynote.

However, the release of Llama 4 has brought to light some systematic issues with LM Arena. The Llama team reported that their "experimental chat version" scored an ELO of 1417 on LM Arena, but the actual model released scored much lower, around 1,200. This discrepancy has led to the publication of the "Leaderboard Illusion" paper, which highlights several concerns with the LM Arena platform.

The paper alleges that there are undisclosed private testing practices that benefit a handful of providers, who are able to test multiple variants before public release and selectively disclose the best-performing results. This has led to a biased leaderboard, where proprietary model providers can gain an advantage by overfitting their models to the specific LM Arena benchmark.

Furthermore, the paper suggests that there is a data access disparity, where the model providers are able to use the data shared by the LM Arena team to fine-tune their models and improve their performance on the leaderboard. This has resulted in a situation where even small amounts of data can lead to significant gains in performance on the LM Arena platform.

The issues raised in the "Leaderboard Illusion" paper are not limited to LM Arena, as similar concerns have been raised about other benchmarks, such as Frontier Math and ARC AGI. These examples highlight the need for greater transparency and accountability in the development and use of AI benchmarks, to ensure a level playing field and accurate representation of model capabilities.

The Leaderboard Illusion paper has highlighted several systematic issues with the LM Arena benchmark:

1. Data Access Disparity: Proprietary model providers are able to gain access to the data used by LM Arena, which allows them to fine-tune their models specifically for this leaderboard. This can lead to biased scores, as these providers can test multiple variants and choose the one that performs best on the leaderboard.

2. Model Removal: The paper found that 66% of silently removed models from LM Arena were open-source or open-weight models, while proprietary models were not removed at the same rate.

3. Overfitting to the Leaderboard: The ability of proprietary providers to test multiple variants and choose the best-performing one leads to models that are optimized specifically for the LM Arena leaderboard, rather than for real-world performance.

4. Lack of Transparency: The paper alleges that LM Arena has "undisclosed private testing practices" that benefit a handful of providers, leading to a distorted playing field.

5. Misalignment of Metrics: The paper suggests that using human preference as the primary metric for a leaderboard like LM Arena can lead to models that optimize for this metric, rather than for more meaningful real-world capabilities.

Overall, the Leaderboard Illusion paper raises serious concerns about the integrity and fairness of the LM Arena benchmark, and highlights the need for more transparent and rigorous evaluation practices in the LLM community.

The paper highlights systematic issues with benchmarks, particularly in the LM (Large Language Model) arena. One of the key problems identified is the data access disparity between proprietary model providers and open-source models.

The authors found that proprietary providers are able to test multiple variants of their models in private before public release, and can selectively disclose the best-performing scores. This allows them to "overfit" their models to the specific LM Arena leaderboard, leading to biased and distorted results.

Specifically, the paper states that these providers were able to test 27 different private models or variants, and then choose the one with the best performance to report on the leaderboard. This gives them a significant advantage over open-source models, which do not have the same level of access to the data and testing capabilities.

Furthermore, the authors found that 66% of the models silently removed from the leaderboard were open-source, while proprietary models were not removed at the same rate. This further exacerbates the uneven playing field.

The paper also highlights the issue of data sharing, where the LM Arena team shares 20% of the data back with the model providers to help them improve their models. This feedback loop allows the proprietary providers to fine-tune their models specifically for the LM Arena leaderboard, leading to a distorted representation of their real-world capabilities.

Overall, the paper argues that these systematic issues have resulted in a "leaderboard illusion," where the rankings do not accurately reflect the true capabilities of the AI systems, but rather the ability of certain providers to game the system.

The paper highlights a concerning issue regarding the removal of models from the LM Arena leaderboard. According to the authors, 66% of the models that were silently removed from the leaderboard were open-source or open-weight models, while the proprietary models were not removed at a similar scale.

This selective removal of models creates a biased playing field, where the open-source and smaller providers are disproportionately affected. The paper argues that the ability of the proprietary model providers to choose the best-performing scores and retain them on the leaderboard leads to distorted and biased arena scores.

This model removal bias undermines the fairness and transparency of the LM Arena leaderboard, as it gives an unfair advantage to the larger, proprietary model providers who can selectively disclose their performance results. The authors suggest that this practice goes against the principles of scientific rigor and transparency that should underpin such benchmarking platforms.

The paper highlights a critical issue with the current state of language model benchmarking - the tendency to optimize for the wrong metrics. The authors argue that the LM Arena leaderboard, which has become an industry standard, suffers from systematic issues that have led to a distorted playing field.

One of the key problems is the ability of proprietary model providers to engage in undisclosed private testing practices. These providers can test multiple variants of their models and selectively disclose the best-performing ones, leading to biased arena scores. This practice of "selective disclosure" allows them to overfit their models to the specific leaderboard, rather than focusing on real-world performance.

Furthermore, the authors point out that the data-sharing practices of the LM Arena team, where they provide 20% of the data back to the model providers, have been exploited by these providers. They can use this data to fine-tune their models and gain an unfair advantage on the leaderboard.

The paper also raises concerns about the removal of models from the leaderboard, where open-source models are disproportionately affected compared to proprietary ones. This further skews the playing field and undermines the transparency and fairness of the benchmark.

The broader implication is that optimizing for the wrong metrics, such as human preference scores on the LM Arena, can lead to the development of models that excel at gaming the system rather than providing genuine value. As highlighted by experts like Alex Albon and Andre Karpathy, this can result in a "toxic feedback loop" where the industry becomes obsessed with chasing better scores on these leaderboards, rather than focusing on building models that truly serve the needs of users.

The paper's findings serve as a wake-up call for the AI community to critically examine the practices and incentives underlying benchmark development and model evaluation. It emphasizes the need for more transparent and rigorous benchmarking approaches that align with real-world performance and value, rather than narrow, gameable metrics.

The paper "The Leaderboard Illusion" has sparked a significant response from the AI community. Several prominent figures have shared their thoughts on the issues raised in the paper.

André Karpathy, a respected AI researcher, expressed suspicion about certain models scoring highly on the LM Arena leaderboard, noting that when he tried to use them, the performance was worse than expected. He also pointed out that the Cloud models, which are highly specialized for coding tasks, have consistently ranked low on the leaderboard, suggesting that the leaderboard may not be a reliable indicator of real-world performance.

Karpathy also highlighted the potential for "gaming the system" by generating large amounts of text with features like bullet points and emojis, which may be favored by the human preference-based scoring used in the LM Arena.

In response to the paper, the LM Arena team has acknowledged some of the concerns raised, but also defended their practices. They argue that the leaderboard reflects the preferences of millions of real users, and that pre-release testing and data sharing with model providers can be beneficial for the community.

However, the LM Arena team also acknowledges that they are working on statistical methods to better understand the components of human preferences, suggesting that they recognize the need to address the issues highlighted in the paper.

The team also disputes some of the specific claims made in the paper, arguing that there are factual errors and misleading statements. They point to a recent blog post that provides more detailed statistics on the number of models tested by different providers.

Overall, the responses from the community and the LM Arena team suggest that there are valid concerns about the transparency and fairness of the leaderboard, but also a recognition that addressing these issues is important for the continued development and deployment of large language models.

The paper on the "leaderboard illusion" has raised significant concerns about the systematic issues plaguing the LLM community, particularly with regards to benchmarks like LM Arena. The key issues highlighted include:

1. Data Access Disparity: Proprietary model providers have access to the data shared by the LM Arena platform, which they can leverage to fine-tune and optimize their models specifically for this leaderboard, leading to biased results.

2. Selective Disclosure of Performance: These providers can test multiple variants of their models and selectively disclose the best-performing ones, creating an illusion of superior capabilities.

3. Model Removal Bias: Open-source models are more likely to be silently removed from the leaderboard compared to proprietary models.

The response from the LM Arena team acknowledges some of these concerns, but also defends their practices, arguing that pre-release testing and optimizing for human preferences are beneficial. However, the community has raised valid criticisms about the potential for abuse and the need for greater transparency.

This issue is not limited to LM Arena; similar concerns have been raised about other benchmarks, such as Frontier Math and ARC AGI, where proprietary access and potential data misuse have been highlighted.

Moving forward, it is crucial for the LLM community to address these systematic issues and work towards more transparent, fair, and scientifically rigorous benchmarking practices. This may involve developing alternative leaderboards, internal benchmarks, and a greater emphasis on real-world performance rather than just human preference scores.