HeadRank: Decoding-Free Passage Reranking via Preference-Aligned Attention Heads

TL;DR

HeadRank introduces a decoding-free passage reranking method that leverages attention weights for efficient and highly accurate document ranking across multiple benchmarks, outperforming existing baselines.

Contribution

The paper presents a novel attention-based reranking framework that enhances preference discrimination in LLM attention weights, reducing inference complexity and improving ranking performance.

Findings

Achieves highest average NDCG@10 across 14 benchmarks on three Qwen3 scales.

Significantly improves middle-zone document relevance discrimination.

Demonstrates 43-percentage-point gap in relevance for top-ranked middle documents.

Abstract

Decoding-free reranking methods that read relevance signals directly from LLM attention weights offer significant latency advantages over autoregressive approaches, yet suffer from attention score homogenization: middle-context documents receive near-identical scores, destroying the fine-grained distinctions required for ranking. We propose HeadRank, a framework that lifts preference optimization from discrete token space into the continuous attention domain through entropy-regularized head selection, hard adjacent-level preference pairs, and a distribution regularizer that jointly sharpen discriminability in the homogenized middle zone. Depth truncation at the deepest selected layer further reduces inference to $\mathcal{O}(1)$ forward passes. Across 14 benchmarks on three Qwen3 scales (0.6B--4B) using only 211 training queries, HeadRank achieves the highest average NDCG@10 at every scale, outperforming both generative and decoding-free baselines on the majority of benchmarks with 100\% formatting success. At 4B, 57.4\% of relevant middle-zone documents reach the top quartile versus 14.2\% for irrelevant ones -- a 43-percentage-point selectivity gap that demonstrates the effectiveness of attention-space preference alignment for listwise reranking.