{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2026:JR6QTSCBVG3MGRCOOAJENO2K74","short_pith_number":"pith:JR6QTSCB","schema_version":"1.0","canonical_sha256":"4c7d09c841a9b6c3444e701246bb4aff03013b4df8f57af1a37a78308522a894","source":{"kind":"arxiv","id":"2606.11198","version":1},"attestation_state":"computed","paper":{"title":"The Structural Attention Tax: How Retrieval Format Hijacks In-Context Learning Independent of Content","license":"http://creativecommons.org/licenses/by/4.0/","headline":"","cross_cats":["cs.AI"],"primary_cat":"cs.CL","authors_text":"Di Zhang, Yuqi Zhang","submitted_at":"2026-04-21T08:47:00Z","abstract_excerpt":"Retrieval-augmented generation (RAG) systems inject external knowledge to improve LLM outputs, yet the format of injected content -- distinct from its semantic relevance -- can independently distort the model's attention distribution. We identify and formalise a phenomenon we term the structural attention tax: knowledge graph (KG) triples, due to their relational delimiters and repeated slot patterns, capture 2-3x more attention per token than semantically equivalent natural-language text ($\\hat{o}$(KG) $\\approx$ 0.70 vs. $\\hat{o}$(neutral) $\\approx$ 0.25), compressing demonstration attention "},"verification_status":{"content_addressed":true,"pith_receipt":true,"author_attested":false,"weak_author_claims":0,"strong_author_claims":0,"externally_anchored":false,"storage_verified":false,"citation_signatures":0,"replication_records":0,"graph_snapshot":true,"references_resolved":false,"formal_links_present":false},"canonical_record":{"source":{"id":"2606.11198","kind":"arxiv","version":1},"metadata":{"license":"http://creativecommons.org/licenses/by/4.0/","primary_cat":"cs.CL","submitted_at":"2026-04-21T08:47:00Z","cross_cats_sorted":["cs.AI"],"title_canon_sha256":"9f54c0415066f2924e810a13a5e3239de7f240508d1f134b8c15b55def183222","abstract_canon_sha256":"08b68db9e69f1eb3bb646f3926e6c6221cd6e2b346b5c7912bfc8535a7a17444"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-06-11T00:08:05.943179Z","signature_b64":"os+oPw9+lXf9e8xBVIMaqcYOg4YgJ4LJK8AApCrqARNKQkUngamRuudZB5JOK5BcvKCsUz7TqgVDNX62/c0ACA==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"4c7d09c841a9b6c3444e701246bb4aff03013b4df8f57af1a37a78308522a894","last_reissued_at":"2026-06-11T00:08:05.942110Z","signature_status":"signed_v1","first_computed_at":"2026-06-11T00:08:05.942110Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"The Structural Attention Tax: How Retrieval Format Hijacks In-Context Learning Independent of Content","license":"http://creativecommons.org/licenses/by/4.0/","headline":"","cross_cats":["cs.AI"],"primary_cat":"cs.CL","authors_text":"Di Zhang, Yuqi Zhang","submitted_at":"2026-04-21T08:47:00Z","abstract_excerpt":"Retrieval-augmented generation (RAG) systems inject external knowledge to improve LLM outputs, yet the format of injected content -- distinct from its semantic relevance -- can independently distort the model's attention distribution. We identify and formalise a phenomenon we term the structural attention tax: knowledge graph (KG) triples, due to their relational delimiters and repeated slot patterns, capture 2-3x more attention per token than semantically equivalent natural-language text ($\\hat{o}$(KG) $\\approx$ 0.70 vs. $\\hat{o}$(neutral) $\\approx$ 0.25), compressing demonstration attention "},"claims":{"count":0,"items":[],"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"source":{"id":"2606.11198","kind":"arxiv","version":1},"verdict":{"id":null,"model_set":{},"created_at":null,"strongest_claim":"","one_line_summary":"","pipeline_version":null,"weakest_assumption":"","pith_extraction_headline":""},"integrity":{"clean":true,"summary":{"advisory":0,"critical":0,"by_detector":{},"informational":0},"endpoint":"/pith/2606.11198/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"},"aliases":[{"alias_kind":"arxiv","alias_value":"2606.11198","created_at":"2026-06-11T00:08:05.942256+00:00"},{"alias_kind":"arxiv_version","alias_value":"2606.11198v1","created_at":"2026-06-11T00:08:05.942256+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2606.11198","created_at":"2026-06-11T00:08:05.942256+00:00"},{"alias_kind":"pith_short_12","alias_value":"JR6QTSCBVG3M","created_at":"2026-06-11T00:08:05.942256+00:00"},{"alias_kind":"pith_short_16","alias_value":"JR6QTSCBVG3MGRCO","created_at":"2026-06-11T00:08:05.942256+00:00"},{"alias_kind":"pith_short_8","alias_value":"JR6QTSCB","created_at":"2026-06-11T00:08:05.942256+00:00"}],"events":[],"event_summary":{},"paper_claims":[],"inbound_citations":{"count":1,"internal_anchor_count":1,"sample":[{"citing_arxiv_id":"2608.09093","citing_title":"The Announcement Carries the Cue: Markup, Boundaries, and the Notation of Pre-Training Corpora","ref_index":91,"is_internal_anchor":true}]},"formal_canon":{"evidence_count":0,"sample":[],"anchors":[]},"links":{"html":"https://pith.science/pith/JR6QTSCBVG3MGRCOOAJENO2K74","json":"https://pith.science/pith/JR6QTSCBVG3MGRCOOAJENO2K74.json","graph_json":"https://pith.science/api/pith-number/JR6QTSCBVG3MGRCOOAJENO2K74/graph.json","events_json":"https://pith.science/api/pith-number/JR6QTSCBVG3MGRCOOAJENO2K74/events.json","paper":"https://pith.science/paper/JR6QTSCB"},"agent_actions":{"view_html":"https://pith.science/pith/JR6QTSCBVG3MGRCOOAJENO2K74","download_json":"https://pith.science/pith/JR6QTSCBVG3MGRCOOAJENO2K74.json","view_paper":"https://pith.science/paper/JR6QTSCB","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2606.11198&json=true","fetch_graph":"https://pith.science/api/pith-number/JR6QTSCBVG3MGRCOOAJENO2K74/graph.json","fetch_events":"https://pith.science/api/pith-number/JR6QTSCBVG3MGRCOOAJENO2K74/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/JR6QTSCBVG3MGRCOOAJENO2K74/action/timestamp_anchor","attest_storage":"https://pith.science/pith/JR6QTSCBVG3MGRCOOAJENO2K74/action/storage_attestation","attest_author":"https://pith.science/pith/JR6QTSCBVG3MGRCOOAJENO2K74/action/author_attestation","sign_citation":"https://pith.science/pith/JR6QTSCBVG3MGRCOOAJENO2K74/action/citation_signature","submit_replication":"https://pith.science/pith/JR6QTSCBVG3MGRCOOAJENO2K74/action/replication_record"}},"created_at":"2026-06-11T00:08:05.942256+00:00","updated_at":"2026-06-11T00:08:05.942256+00:00"}