Are we overfitting to benchmarks?

I’ve been noticing a worrying trend in the latest LLM releases. Performance on popular benchmarks keeps climbing, but real-world reasoning improvements seem marginal at best. Some papers now show that models trained on benchmark-heavy data can game the metrics without genuine generalization. Are we hitting a point where our evaluation methods are actively steering research away from robust capabilities? I’d love to hear if anyone has recent findings on more adaptive or adversarial evaluation frameworks. Also, how do we balance standardized testing with reproducible, practical assessment? Maybe we need to rethink what we’re optimizing for entirely.