Expert Warns: Kaggle Public Leaderboard Scores Prone to Overfitting

calabi_and_yau · x · 2026-08-13

In response to recent claims that the AI model 'Locus' outperformed 89.5% of human competitors across multiple Kaggle competitions, expert Jean-François Puget raised strong skepticism.

He pointed out that the ongoing competitions' current scores are solely based on the public leaderboard. Climbing the public LB can often be achieved through aggressive overfitting, which does not reflect the model's actual performance on the final hidden test data. Furthermore, the claimed '4th place average' is misleading because it only looks at the few humans who enter all competitions, who are typically not the top contestants.

Original post →

More from Models

Models channel →