I also faced LB overfitting with ensembling. The key is diversity and validation alignment. I used models with different architectures (e.g., XGBoost, CatBoost, NN) trained on varied features/folds, and weighted them based on out-of-fold CV AUC (not LB probing). Also, I capped ensemble size — beyond 3–4 diverse models, gains were negligible and overfitting risk increased. TPU on Kaggle helped run fast experiments. My CV and LB eventually aligned when I ensured temporal split matched data distribution