Learning-based cache eviction algorithms, particularly object-level approaches, have demonstrated superior reduction in miss ratio by making fine-grained eviction decisions. However, this accuracy often comes at a high computational cost, as existing methods typically apply complex prediction models uniformly to all objects and struggle to quickly differentiate hot from cold items. In this paper, we identify two critical insights: (1) not all objects necessitate precise prediction, especially one-hit wonders lacking rich historical features, and (2) effective sampling must rapidly protect hot objects and identify cold eviction candidates. Based on these insights, we propose SL-Cache (Selective Learning Cache), a novel object-level learning eviction algorithm that optimally balances miss ratio and throughput. SL-Cache introduces two key innovations: first, it bypasses costly model predictions for low-utility one-hit wonders by employing precomputed utility estimates, substantially reducing computational overhead. Second, it implements an adaptive, hotness-aware sampling strategy that dynamically demotes objects based on refined access patterns, ensuring hot objects are retained while non-hot candidates are efficiently selected for prediction or eviction. Our extensive experimental evaluation across ten state-of-the-art eviction algorithms, using diverse production and benchmark traces, demonstrates that SL-Cache consistently achieves a superior balance. Specifically, SL-Cache reduces the miss ratio by 11.90% relative to LRU, which is 1.1x higher than the state-of-the-art 3L-Cache, while simultaneously delivering 2x the throughput of 3L-Cache. Furthermore, SL-Cache achieves the highest throughput among all object-level learning algorithms.