“Open leaderboards like the OpenLLM Leaderboard are susceptible to models overfitting the benchmarks, which may not reflect true general capabilities.”