Despite the success of neural models in many major machine learning problems and recently published neural learning to rank (LTR) papers in top venues, the effectiveness of neural models on traditional LTR problems is still not widely acknowledged. We first validate the concern by showing that most recent neural LTR models are, by a large margin, inferior to the best publicly available tree-based implementation, which is sometimes ignored in recent neural LTR papers. We then investigate why existing neural LTR suffers by identifying several of its weaknesses. To that end, we propose a new neural LTR framework that mitigates these weaknesses, by borrowing ideas from several research fields. Our models are able to perform comparatively with the strong tree-based baseline, while outperforming recently published neural learning to rank methods by a large margin. Our results also serve as a benchmark for neural learning to rank models.