{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2025:EDEWANGTS6SAPNY35BFXSTNGVB","short_pith_number":"pith:EDEWANGT","schema_version":"1.0","canonical_sha256":"20c96034d397a407b71be84b794da6a84c95d0668a68178abe7caeb6cb03dfe0","source":{"kind":"arxiv","id":"2508.02546","version":1},"attestation_state":"computed","paper":{"title":"What are you sinking? A geometric approach on attention sink","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","headline":"","cross_cats":["cs.AI","cs.CL"],"primary_cat":"cs.LG","authors_text":"Fabrizio Silvestri, Umberto Nanni, Valeria Ruscio","submitted_at":"2025-08-04T15:59:15Z","abstract_excerpt":"Attention sink (AS) is a consistent pattern in transformer attention maps where certain tokens (often special tokens or positional anchors) disproportionately attract attention from other tokens. We show that in transformers, AS is not an architectural artifact, but it is the manifestation of a fundamental geometric principle: the establishment of reference frames that anchor representational spaces. We analyze several architectures and identify three distinct reference frame types, centralized, distributed, and bidirectional, that correlate with the attention sink phenomenon. We show that the"},"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":"2508.02546","kind":"arxiv","version":1},"metadata":{"license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","primary_cat":"cs.LG","submitted_at":"2025-08-04T15:59:15Z","cross_cats_sorted":["cs.AI","cs.CL"],"title_canon_sha256":"342a54ec1525a380fd3e1ba6463675be84772f24a181c210b5ec809585e52339","abstract_canon_sha256":"da27ead97f719538c11ecb992c282c84d890c96f743c2093d7f183ea8136ab23"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-07-05T11:48:16.040730Z","signature_b64":"FELblDaZX6LFmLlOcs7Pt/FfTJwYol/OGa/MYtNU1ZJBOxMVrSE0twOsEJWgL0YSgxmGS9dNreaaQltdFtUrBw==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"20c96034d397a407b71be84b794da6a84c95d0668a68178abe7caeb6cb03dfe0","last_reissued_at":"2026-07-05T11:48:16.040221Z","signature_status":"signed_v1","first_computed_at":"2026-07-05T11:48:16.040221Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"What are you sinking? A geometric approach on attention sink","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","headline":"","cross_cats":["cs.AI","cs.CL"],"primary_cat":"cs.LG","authors_text":"Fabrizio Silvestri, Umberto Nanni, Valeria Ruscio","submitted_at":"2025-08-04T15:59:15Z","abstract_excerpt":"Attention sink (AS) is a consistent pattern in transformer attention maps where certain tokens (often special tokens or positional anchors) disproportionately attract attention from other tokens. We show that in transformers, AS is not an architectural artifact, but it is the manifestation of a fundamental geometric principle: the establishment of reference frames that anchor representational spaces. We analyze several architectures and identify three distinct reference frame types, centralized, distributed, and bidirectional, that correlate with the attention sink phenomenon. We show that the"},"claims":{"count":0,"items":[],"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"source":{"id":"2508.02546","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/2508.02546/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":"2508.02546","created_at":"2026-07-05T11:48:16.040280+00:00"},{"alias_kind":"arxiv_version","alias_value":"2508.02546v1","created_at":"2026-07-05T11:48:16.040280+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2508.02546","created_at":"2026-07-05T11:48:16.040280+00:00"},{"alias_kind":"pith_short_12","alias_value":"EDEWANGTS6SA","created_at":"2026-07-05T11:48:16.040280+00:00"},{"alias_kind":"pith_short_16","alias_value":"EDEWANGTS6SAPNY3","created_at":"2026-07-05T11:48:16.040280+00:00"},{"alias_kind":"pith_short_8","alias_value":"EDEWANGT","created_at":"2026-07-05T11:48:16.040280+00:00"}],"events":[],"event_summary":{},"paper_claims":[],"inbound_citations":{"count":11,"internal_anchor_count":0,"sample":[{"citing_arxiv_id":"2606.02288","citing_title":"Massive Spikes in LLMs are Bias Vectors: Mechanistic Uncovering and Spike-Free Quantization","ref_index":2,"is_internal_anchor":false},{"citing_arxiv_id":"2605.16331","citing_title":"Retrieval and competition: how a protein foundation model starts a protein","ref_index":16,"is_internal_anchor":false},{"citing_arxiv_id":"2605.22372","citing_title":"ASAP: Attention Sink Anchored Pruning","ref_index":16,"is_internal_anchor":false},{"citing_arxiv_id":"2604.10027","citing_title":"SinkTrack: Attention Sink based Context Anchoring for Large Language Models","ref_index":13,"is_internal_anchor":false},{"citing_arxiv_id":"2605.16331","citing_title":"Retrieval and competition: how a protein foundation model starts a protein","ref_index":16,"is_internal_anchor":false},{"citing_arxiv_id":"2605.08504","citing_title":"A Single Layer to Explain Them All:Understanding Massive Activations in Large Language Models","ref_index":17,"is_internal_anchor":false},{"citing_arxiv_id":"2604.03316","citing_title":"When Sinks Help or Hurt: Unified Framework for Attention Sink in Large Vision-Language Models","ref_index":32,"is_internal_anchor":false},{"citing_arxiv_id":"2604.22783","citing_title":"Parameter Efficiency Is Not Memory Efficiency: Rethinking Fine-Tuning for On-Device LLM Adaptation","ref_index":3,"is_internal_anchor":false},{"citing_arxiv_id":"2605.08504","citing_title":"A Single Layer to Explain Them All:Understanding Massive Activations in Large Language Models","ref_index":17,"is_internal_anchor":false},{"citing_arxiv_id":"2604.10027","citing_title":"SinkTrack: Attention Sink based Context Anchoring for Large Language Models","ref_index":13,"is_internal_anchor":false},{"citing_arxiv_id":"2604.14722","citing_title":"A Mechanistic Account of Attention Sinks in GPT-2: One Circuit, Broader Implications for Mitigation","ref_index":4,"is_internal_anchor":false}]},"formal_canon":{"evidence_count":0,"sample":[],"anchors":[]},"links":{"html":"https://pith.science/pith/EDEWANGTS6SAPNY35BFXSTNGVB","json":"https://pith.science/pith/EDEWANGTS6SAPNY35BFXSTNGVB.json","graph_json":"https://pith.science/api/pith-number/EDEWANGTS6SAPNY35BFXSTNGVB/graph.json","events_json":"https://pith.science/api/pith-number/EDEWANGTS6SAPNY35BFXSTNGVB/events.json","paper":"https://pith.science/paper/EDEWANGT"},"agent_actions":{"view_html":"https://pith.science/pith/EDEWANGTS6SAPNY35BFXSTNGVB","download_json":"https://pith.science/pith/EDEWANGTS6SAPNY35BFXSTNGVB.json","view_paper":"https://pith.science/paper/EDEWANGT","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2508.02546&json=true","fetch_graph":"https://pith.science/api/pith-number/EDEWANGTS6SAPNY35BFXSTNGVB/graph.json","fetch_events":"https://pith.science/api/pith-number/EDEWANGTS6SAPNY35BFXSTNGVB/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/EDEWANGTS6SAPNY35BFXSTNGVB/action/timestamp_anchor","attest_storage":"https://pith.science/pith/EDEWANGTS6SAPNY35BFXSTNGVB/action/storage_attestation","attest_author":"https://pith.science/pith/EDEWANGTS6SAPNY35BFXSTNGVB/action/author_attestation","sign_citation":"https://pith.science/pith/EDEWANGTS6SAPNY35BFXSTNGVB/action/citation_signature","submit_replication":"https://pith.science/pith/EDEWANGTS6SAPNY35BFXSTNGVB/action/replication_record"}},"created_at":"2026-07-05T11:48:16.040280+00:00","updated_at":"2026-07-05T11:48:16.040280+00:00"}