Meta faces turmoil over alleged Llama 4 benchmark tampering and resignations

Benchmark scores keep climbing, but real-world model performance seems to tell a very different story.

Whistleblower alleges benchmark cheating on Llama 4

According to a report shared by tech blogger Tony Peng, a Chinese engineer on Meta's AI team has resigned and explicitly requested to be removed from the upcoming Llama 4 technical report.

The whistleblower claims that despite repeated training runs, the internal model still lagged far behind open-source state-of-the-art levels. With an internal late-April deadline looming to justify continued investment, leadership allegedly suggested contaminating the post-training process with benchmark test sets to produce passable metrics.

Following the model's release yesterday, negative user feedback surged across Reddit and X. The whistleblower noted that as an active academic researcher, such practices crossed an ethical line, adding that Meta's VP of AI reportedly stepped down for the same reason.

Why it matters

If these allegations hold up, it reveals how extreme benchmark anxiety has become inside Big Tech's open-source race.

Gaming evaluations during post-training is essentially cheating on the test. While it produces impressive leaderboards to secure budget, the deception collapses immediately under real-world developer workloads. We think teams should hold off on relying on Llama 4 until independent community evaluations establish its actual capabilities.

It also highlights the rapid erosion of trust in standard AI benchmarks, proving once again that synthetic metrics are increasingly useless without uncorrupted evaluation.