HeadRank: Decoding-Free Passage Reranking via Preference-Aligned Attention Heads
For practitioners of LLM-based retrieval, HeadRank provides a fast, training-efficient reranking method that overcomes a key limitation of decoding-free approaches.
HeadRank addresses attention score homogenization in decoding-free passage reranking by applying preference optimization directly to attention weights, achieving a 43-percentage-point selectivity gap between relevant and irrelevant middle-context documents at 4B scale, outperforming baselines across 14 benchmarks with only 211 training queries.
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 consistently outperforms generative and decoding-free baselines 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.