Yann LeCun Says Llama 4 Benchmark Results Were Fudged a Little Bit

Llama 4 Maverick's experimental build scored an Arena Elo of 1417, then fell to 32nd when the released model was retested; Yann LeCun says Meta's results were "fudged a little bit." Contrary Research argues such benchmarks steer reinforcement learning toward wrong targets as labs spend $700 billion on 2026 compute.

• A paper found Meta privately tested 27 unreleased Llama 4 variants on Arena before launch, per the Leaderboard Illusion researchers.
• Mark Zuckerberg says the released pure model has no tuning for the leaderboard, calling the gap a reason to be careful with benchmarks.
• Arena still lets labs test uncapped unreleased variants and withdraw any model version they dislike, 18 months after Maverick.