{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2024:TCLJGUJEME4HUOZWGLRGCI7Q56","short_pith_number":"pith:TCLJGUJE","schema_version":"1.0","canonical_sha256":"989693512461387a3b3632e26123f0ef8a1e5b886b9c6a9de6c0f345e42bf92e","source":{"kind":"arxiv","id":"2410.13835","version":2},"attestation_state":"computed","paper":{"title":"Active-Dormant Attention Heads: Mechanistically Demystifying Extreme-Token Phenomena in LLMs","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","headline":"","cross_cats":[],"primary_cat":"cs.LG","authors_text":"Druv Pai, Jiantao Jiao, Michael I. Jordan, Song Mei, Tianyu Guo, Yu Bai","submitted_at":"2024-10-17T17:54:06Z","abstract_excerpt":"Practitioners have consistently observed three puzzling phenomena in transformer-based large language models (LLMs): attention sinks, value-state drains, and residual-state peaks, collectively referred to as extreme-token phenomena. These phenomena are characterized by certain so-called \"sink tokens\" receiving disproportionately high attention weights, exhibiting significantly smaller value states, and having much larger residual-state norms than those of other tokens. These extreme tokens give rise to various challenges in LLM inference, quantization, and interpretability.\n  We elucidate 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":"2410.13835","kind":"arxiv","version":2},"metadata":{"license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","primary_cat":"cs.LG","submitted_at":"2024-10-17T17:54:06Z","cross_cats_sorted":[],"title_canon_sha256":"4c4d621b4d3e8f9f05007e80168a0021a5ebcd7653c5ed71ddd2fb50325edad6","abstract_canon_sha256":"29e149fd043bfc1355dfe8500066edcd47d8e9aff363e8fa05a63ce975fb6727"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-07-05T09:32:33.084504Z","signature_b64":"I8/xWeZvgPDD3CEdDF7SGBTh9miFLcTurDA0s9ST9sYio4NWbpuPpA92zTw4P8Con8O/jklfLuT9j6yFaCWtAQ==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"989693512461387a3b3632e26123f0ef8a1e5b886b9c6a9de6c0f345e42bf92e","last_reissued_at":"2026-07-05T09:32:33.083880Z","signature_status":"signed_v1","first_computed_at":"2026-07-05T09:32:33.083880Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"Active-Dormant Attention Heads: Mechanistically Demystifying Extreme-Token Phenomena in LLMs","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","headline":"","cross_cats":[],"primary_cat":"cs.LG","authors_text":"Druv Pai, Jiantao Jiao, Michael I. Jordan, Song Mei, Tianyu Guo, Yu Bai","submitted_at":"2024-10-17T17:54:06Z","abstract_excerpt":"Practitioners have consistently observed three puzzling phenomena in transformer-based large language models (LLMs): attention sinks, value-state drains, and residual-state peaks, collectively referred to as extreme-token phenomena. These phenomena are characterized by certain so-called \"sink tokens\" receiving disproportionately high attention weights, exhibiting significantly smaller value states, and having much larger residual-state norms than those of other tokens. These extreme tokens give rise to various challenges in LLM inference, quantization, and interpretability.\n  We elucidate the "},"claims":{"count":0,"items":[],"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"source":{"id":"2410.13835","kind":"arxiv","version":2},"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/2410.13835/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":"2410.13835","created_at":"2026-07-05T09:32:33.083934+00:00"},{"alias_kind":"arxiv_version","alias_value":"2410.13835v2","created_at":"2026-07-05T09:32:33.083934+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2410.13835","created_at":"2026-07-05T09:32:33.083934+00:00"},{"alias_kind":"pith_short_12","alias_value":"TCLJGUJEME4H","created_at":"2026-07-05T09:32:33.083934+00:00"},{"alias_kind":"pith_short_16","alias_value":"TCLJGUJEME4HUOZW","created_at":"2026-07-05T09:32:33.083934+00:00"},{"alias_kind":"pith_short_8","alias_value":"TCLJGUJE","created_at":"2026-07-05T09:32:33.083934+00:00"}],"events":[],"event_summary":{},"paper_claims":[],"inbound_citations":{"count":8,"internal_anchor_count":0,"sample":[{"citing_arxiv_id":"2605.28600","citing_title":"Transformers Provably Learn to Internalize Chain-of-Thought","ref_index":13,"is_internal_anchor":false},{"citing_arxiv_id":"2606.07604","citing_title":"Contribution Weights: A Geometrical Analysis of Self-Attention Transformers","ref_index":160,"is_internal_anchor":false},{"citing_arxiv_id":"2606.07604","citing_title":"Contribution Weights: A Geometrical Analysis of Self-Attention Transformers","ref_index":87,"is_internal_anchor":false},{"citing_arxiv_id":"2605.19660","citing_title":"OScaR: The Occam's Razor for Extreme KV Cache Quantization in LLMs and Beyond","ref_index":14,"is_internal_anchor":false},{"citing_arxiv_id":"2410.10781","citing_title":"When Attention Sink Emerges in Language Models: An Empirical View","ref_index":19,"is_internal_anchor":false},{"citing_arxiv_id":"2602.01203","citing_title":"Attention Sink Forges Native MoE in Attention Layers: Sink-Aware Training to Address Head Collapse","ref_index":16,"is_internal_anchor":false},{"citing_arxiv_id":"2604.10098","citing_title":"Attention Sink in Transformers: A Survey on Utilization, Interpretation, and Mitigation","ref_index":163,"is_internal_anchor":false},{"citing_arxiv_id":"2604.14922","citing_title":"LongAct: Harnessing Intrinsic Activation Patterns for Long-Context Reinforcement Learning","ref_index":11,"is_internal_anchor":false}]},"formal_canon":{"evidence_count":0,"sample":[],"anchors":[]},"links":{"html":"https://pith.science/pith/TCLJGUJEME4HUOZWGLRGCI7Q56","json":"https://pith.science/pith/TCLJGUJEME4HUOZWGLRGCI7Q56.json","graph_json":"https://pith.science/api/pith-number/TCLJGUJEME4HUOZWGLRGCI7Q56/graph.json","events_json":"https://pith.science/api/pith-number/TCLJGUJEME4HUOZWGLRGCI7Q56/events.json","paper":"https://pith.science/paper/TCLJGUJE"},"agent_actions":{"view_html":"https://pith.science/pith/TCLJGUJEME4HUOZWGLRGCI7Q56","download_json":"https://pith.science/pith/TCLJGUJEME4HUOZWGLRGCI7Q56.json","view_paper":"https://pith.science/paper/TCLJGUJE","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2410.13835&json=true","fetch_graph":"https://pith.science/api/pith-number/TCLJGUJEME4HUOZWGLRGCI7Q56/graph.json","fetch_events":"https://pith.science/api/pith-number/TCLJGUJEME4HUOZWGLRGCI7Q56/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/TCLJGUJEME4HUOZWGLRGCI7Q56/action/timestamp_anchor","attest_storage":"https://pith.science/pith/TCLJGUJEME4HUOZWGLRGCI7Q56/action/storage_attestation","attest_author":"https://pith.science/pith/TCLJGUJEME4HUOZWGLRGCI7Q56/action/author_attestation","sign_citation":"https://pith.science/pith/TCLJGUJEME4HUOZWGLRGCI7Q56/action/citation_signature","submit_replication":"https://pith.science/pith/TCLJGUJEME4HUOZWGLRGCI7Q56/action/replication_record"}},"created_at":"2026-07-05T09:32:33.083934+00:00","updated_at":"2026-07-05T09:32:33.083934+00:00"}