Meta is facing criticism after it was revealed the company used a customized version of its Llama 4 Maverick model to achieve a high ranking on the popular AI benchmarking site LMArena. The model, Maverick, was promoted as surpassing OpenAI‘s GPT-4o and Google’s Gemini 2.0 Flash, earning a strong ELO score that briefly positioned it as a top contender. However, researchers discovered that the version tested wasn’t the same as the publicly released one. Meta had submitted an “experimental chat version” optimized specifically for conversational tasks to LMArena without clearly disclosing this detail.
LMArena responded by updating its policies to prevent such incidents in the future, stating that Meta’s approach didn’t align with expectations for transparency. A Meta spokesperson explained the company often experiments with model variants and has since released the open-source version of Llama 4. The company denied allegations that it trained the model using test set data, attributing inconsistencies in performance to ongoing development work.
The move raised concerns about benchmark manipulation, with critics arguing that allowing companies to submit tuned models distorts real-world performance insights. Developers who rely on benchmark results to make implementation decisions may end up using versions of models that don’t reflect the advertised performance. The timing of Llama 4’s weekend release and the lack of clarity around the model’s capabilities added to the confusion.
This incident underscores growing tensions in the AI space over how performance is measured and presented. As AI tools become more central to tech innovation, the integrity of benchmark reporting has become a key issue. Meta’s actions, while not in direct violation of LMArena’s rules, have prompted calls for stricter standards to ensure fair, transparent evaluations in the future.
