{"paper":{"title":"HeadRank: Decoding-Free Passage Reranking via Preference-Aligned Attention Heads","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","headline":"HeadRank reranks passages by aligning LLM attention heads to preferences in continuous space without decoding.","cross_cats":["cs.AI"],"primary_cat":"cs.IR","authors_text":"Aolin Li, Chenxing Wang, Dongliang Liao, Haijun Wu, Huiyun Hu, Jin Xu, Junwu Du, Juyuan Wang, Ligang Liu, Shunlin Rong, Yuchen Fang","submitted_at":"2026-04-19T03:43:42Z","abstract_excerpt":"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 dis"},"claims":{"count":4,"items":[{"kind":"strongest_claim","text":"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.","source":"verdict.strongest_claim","status":"machine_extracted","claim_id":"C1","attestation":"unclaimed"},{"kind":"weakest_assumption","text":"That entropy-regularized head selection combined with hard adjacent-level preference pairs and a distribution regularizer can reliably overcome attention homogenization in middle context using only 211 training queries without introducing new biases or overfitting to the small training set.","source":"verdict.weakest_assumption","status":"machine_extracted","claim_id":"C2","attestation":"unclaimed"},{"kind":"one_line_summary","text":"HeadRank improves decoding-free passage reranking by preference-aligning attention heads to increase discriminability in middle-context documents, outperforming baselines on 14 benchmarks with only 211 training queries.","source":"verdict.one_line_summary","status":"machine_extracted","claim_id":"C3","attestation":"unclaimed"},{"kind":"headline","text":"HeadRank reranks passages by aligning LLM attention heads to preferences in continuous space without decoding.","source":"verdict.pith_extraction.headline","status":"machine_extracted","claim_id":"C4","attestation":"unclaimed"}],"snapshot_sha256":"6ba0c05f1268c983e42c044331acbab37d2e608a6661f2f2f869a04e2edd3fb2"},"source":{"id":"2604.17237","kind":"arxiv","version":2},"verdict":{"id":"873dd888-bab8-4e3b-8f88-e5afebace4d0","model_set":{"reader":"grok-4.3"},"created_at":"2026-05-10T06:31:49.425911Z","strongest_claim":"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.","one_line_summary":"HeadRank improves decoding-free passage reranking by preference-aligning attention heads to increase discriminability in middle-context documents, outperforming baselines on 14 benchmarks with only 211 training queries.","pipeline_version":"pith-pipeline@v0.9.0","weakest_assumption":"That entropy-regularized head selection combined with hard adjacent-level preference pairs and a distribution regularizer can reliably overcome attention homogenization in middle context using only 211 training queries without introducing new biases or overfitting to the small training set.","pith_extraction_headline":"HeadRank reranks passages by aligning LLM attention heads to preferences in continuous space without decoding."},"integrity":{"clean":true,"summary":{"advisory":0,"critical":0,"by_detector":{},"informational":0},"endpoint":"/pith/2604.17237/integrity.json","findings":[],"available":true,"detectors_run":[],"snapshot_sha256":"c28c3603d3b5d939e8dc4c7e95fa8dfce3d595e45f758748cecf8e644a296938"},"references":{"count":0,"sample":[],"resolved_work":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57","internal_anchors":0},"formal_canon":{"evidence_count":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"author_claims":{"count":0,"strong_count":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"builder_version":"pith-number-builder-2026-05-17-v1"}