HLATR Hybrid List Aware Transformer Reranking

A lightweight list-aware neural reranking architecture for multi-stage text retrieval that couples coarse first-stage retrieval features with fine-grained second-stage interaction-based reranker representations across a candidate document list D'_r of size Z by mapping each document d_i's first-stage retrieval rank order i (0 <= i <= Z) to a learnable position embedding pe_i in R^d, projecting its second-stage pretrained language model reranker representation E^R(q, d_i) via v_i = W_v^T E^R(q, d_i) in R^d, fusing both into an initial sequence matrix H^0 = [LN(pe_1 + v_1), ..., LN(pe_Z + v_Z)]^T in R^{d x Z}, passing H^0 through L Transformer encoder layers H^l = T^l(H^{l-1}) to capture mutual partial-order dependencies among all Z candidate documents simultaneously, and predicting final relevance scores score(q, d_i) = W^T H_i^l trained with a listwise contrastive loss L_{q, D'_r} = -log(exp(score(q, d+)) / sum_{d in D'_r} exp(score(q, d))).