Recently, Yann LeCun, a trailblazer in artificial intelligence and a recipient of the Turing Award, made a public confirmation for the first time after departing from Meta. He stated that the company's large language model, Llama 4, which was launched in April 2025, resorted to fraudulent practices in key benchmark tests.

He straightforwardly pointed out that the test results were somewhat embellished. The team employed different model variants for various benchmark tests to secure higher scores. This delayed admission not only validated the widespread suspicions within the developer community months prior but also unveiled the deep-rooted strategic disputes, management clashes, and personnel upheavals within Meta's AI division.

The roots of this 'benchmark manipulation' incident can be traced back to the release of Llama 4 in April 2025.

Initially, one of its iterations, Llama-4-Maverick, secured the second spot on the prestigious large model arena (Chatbot Arena) leaderboard, showcasing remarkable performance.

However, developers swiftly noticed substantial disparities between the version accessible in the open-source community and the one evaluated for the leaderboard. The actual performance lagged significantly behind the level suggested by the ranking, especially struggling with tasks like programming.

Subsequently, the leaderboard operator verified that Meta had submitted a specially optimized 'custom version' designed for dialogue and updated the ranking accordingly. After retesting, Llama-4-Maverick's ranking tumbled from the previous 2nd place to 32nd.

At that time, Meta officials promptly refuted allegations of 'training cheating.' Yann LeCun, for the first time, acknowledged the manipulation from the vantage point of a core insider. According to him, Meta's CEO, Mark Zuckerberg, was 'extremely disheartened' and 'lost faith' in the entire generative AI (GenAI) team, subsequently sidelining the entire department. This directly set off a wave of aggressive organizational restructuring and talent purges.

The debacle of Llama 4 and its ensuing scandals acted as a catalyst for the long-festering conflicts within Meta, primarily surfacing in two domains: technical routes and personnel management.

Firstly, there was a fundamental disagreement over technical routes. Yann LeCun has been a vocal critic of the prevailing mainstream large language model (LLM) trajectory. He has repeatedly asserted in public that while LLMs are beneficial for language processing, their essence is constrained by linguistic text and lacks the capacity to comprehend the physical world. Pursuing human-level intelligence (superintelligence) through this path is a 'blind alley'.

He advocates for the development of 'world models' capable of grasping the principles governing the world's functioning. However, in Meta's new round of superintelligence project recruitment, the primary emphasis was on large language model development. Yann LeCun felt his perspectives had become a minority within the company and even bluntly stated that many colleagues did not want him to publicly denigrate the LLM route, rendering his position 'administratively challenging' and ultimately leading to his departure.

Secondly, there were contentious high-level personnel shifts. To regroup, Zuckerberg made a substantial investment in data annotation company Scale AI in June 2025 and recruited its 28-year-old founder, Alexander Wang, to head the newly established 'Superintelligence Lab.' This young executive also briefly became Yann LeCun's superior.

Yann LeCun viewed this arrangement unfavorably, stating that while Alexander Wang was a rapid learner, he was 'young and inexperienced,' 'devoid of scientific research experience, unaware of how to conduct research, and unfamiliar with researchers' preferences.' He emphasized that genuine scientific research cannot be propelled by administrative directives. This generational transition and ideological clash in the management tier intensified organizational instability.

Meanwhile, other teams within Meta were also impacted. Tian Yuandong, the former Director of Research Scientists at FAIR (Fundamental AI Research), mentioned in his personal summary that his team had been temporarily reassigned to 'salvage' the Llama 4 project. However, after completing the Llama 4.5 project, the entire team was laid off.

Now, having left Meta, Yann LeCun has established a new company dedicated to researching his cherished 'world models.' His revelations not only constitute a grave accusation against his former employer but also pose a profound inquiry to the entire industry regarding innovation ethics, technological direction, and corporate governance. In the high-stakes race of artificial intelligence, upholding integrity and foresight while striving for leadership is a challenge all participants must confront.

References:

https://news.futunn.com/post/66892851/after-lecun-left-meta-he-made-shocking-revelations-models-cheating?futusource=news_newspage_recommend&level=2&data_ticket=1766116194545042

https://36kr.com/p/3624814768669955

https://m.ithome.com/html/845134.htm