{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2024:F3WUDZEAMQX5PXM3BPC2ZJNQGX","short_pith_number":"pith:F3WUDZEA","schema_version":"1.0","canonical_sha256":"2eed41e480642fd7dd9b0bc5aca5b035ee0db36e523a24126275b43facb80b30","source":{"kind":"arxiv","id":"2402.09221","version":1},"attestation_state":"computed","paper":{"title":"Spectral Filters, Dark Signals, and Attention Sinks","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","headline":"","cross_cats":["cs.CL"],"primary_cat":"cs.AI","authors_text":"Nicola Cancedda","submitted_at":"2024-02-14T15:01:07Z","abstract_excerpt":"Projecting intermediate representations onto the vocabulary is an increasingly popular interpretation tool for transformer-based LLMs, also known as the logit lens. We propose a quantitative extension to this approach and define spectral filters on intermediate representations based on partitioning the singular vectors of the vocabulary embedding and unembedding matrices into bands. We find that the signals exchanged in the tail end of the spectrum are responsible for attention sinking (Xiao et al. 2023), of which we provide an explanation. We find that the loss of pretrained models can be kep"},"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":"2402.09221","kind":"arxiv","version":1},"metadata":{"license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","primary_cat":"cs.AI","submitted_at":"2024-02-14T15:01:07Z","cross_cats_sorted":["cs.CL"],"title_canon_sha256":"3e069eb900ec2a3e047758a46888c2b1b7cf12cc88fd2d0eb8e3b41488dcb112","abstract_canon_sha256":"7e75d6a3f15f1f077a9fcec6e280985ede1c04811355c594a8e68ff2255bc757"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-07-05T07:45:10.230143Z","signature_b64":"tMHtQK83JjSRUrvuHavUOnRjP4pHHzUDBBQJWEQkEnJIDhJbnZ2/3g0dXnUBGxQYRw98BlkxCxmcgFs6mSgDBw==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"2eed41e480642fd7dd9b0bc5aca5b035ee0db36e523a24126275b43facb80b30","last_reissued_at":"2026-07-05T07:45:10.229742Z","signature_status":"signed_v1","first_computed_at":"2026-07-05T07:45:10.229742Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"Spectral Filters, Dark Signals, and Attention Sinks","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","headline":"","cross_cats":["cs.CL"],"primary_cat":"cs.AI","authors_text":"Nicola Cancedda","submitted_at":"2024-02-14T15:01:07Z","abstract_excerpt":"Projecting intermediate representations onto the vocabulary is an increasingly popular interpretation tool for transformer-based LLMs, also known as the logit lens. We propose a quantitative extension to this approach and define spectral filters on intermediate representations based on partitioning the singular vectors of the vocabulary embedding and unembedding matrices into bands. We find that the signals exchanged in the tail end of the spectrum are responsible for attention sinking (Xiao et al. 2023), of which we provide an explanation. We find that the loss of pretrained models can be kep"},"claims":{"count":0,"items":[],"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"source":{"id":"2402.09221","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/2402.09221/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":"2402.09221","created_at":"2026-07-05T07:45:10.229804+00:00"},{"alias_kind":"arxiv_version","alias_value":"2402.09221v1","created_at":"2026-07-05T07:45:10.229804+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2402.09221","created_at":"2026-07-05T07:45:10.229804+00:00"},{"alias_kind":"pith_short_12","alias_value":"F3WUDZEAMQX5","created_at":"2026-07-05T07:45:10.229804+00:00"},{"alias_kind":"pith_short_16","alias_value":"F3WUDZEAMQX5PXM3","created_at":"2026-07-05T07:45:10.229804+00:00"},{"alias_kind":"pith_short_8","alias_value":"F3WUDZEA","created_at":"2026-07-05T07:45:10.229804+00:00"}],"events":[],"event_summary":{},"paper_claims":[],"inbound_citations":{"count":6,"internal_anchor_count":0,"sample":[{"citing_arxiv_id":"2606.07604","citing_title":"Contribution Weights: A Geometrical Analysis of Self-Attention Transformers","ref_index":148,"is_internal_anchor":false},{"citing_arxiv_id":"2606.07604","citing_title":"Contribution Weights: A Geometrical Analysis of Self-Attention Transformers","ref_index":17,"is_internal_anchor":false},{"citing_arxiv_id":"2410.10781","citing_title":"When Attention Sink Emerges in Language Models: An Empirical View","ref_index":5,"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":3,"is_internal_anchor":false},{"citing_arxiv_id":"2605.10622","citing_title":"Vocabulary Hijacking in LVLMs: Unveiling Critical Attention Heads by Excluding Inert Tokens to Mitigate Hallucination","ref_index":84,"is_internal_anchor":false},{"citing_arxiv_id":"2605.06611","citing_title":"The Structural Origin of Attention Sink: Variance Discrepancy, Super Neurons, and Dimension Disparity","ref_index":4,"is_internal_anchor":false}]},"formal_canon":{"evidence_count":0,"sample":[],"anchors":[]},"links":{"html":"https://pith.science/pith/F3WUDZEAMQX5PXM3BPC2ZJNQGX","json":"https://pith.science/pith/F3WUDZEAMQX5PXM3BPC2ZJNQGX.json","graph_json":"https://pith.science/api/pith-number/F3WUDZEAMQX5PXM3BPC2ZJNQGX/graph.json","events_json":"https://pith.science/api/pith-number/F3WUDZEAMQX5PXM3BPC2ZJNQGX/events.json","paper":"https://pith.science/paper/F3WUDZEA"},"agent_actions":{"view_html":"https://pith.science/pith/F3WUDZEAMQX5PXM3BPC2ZJNQGX","download_json":"https://pith.science/pith/F3WUDZEAMQX5PXM3BPC2ZJNQGX.json","view_paper":"https://pith.science/paper/F3WUDZEA","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2402.09221&json=true","fetch_graph":"https://pith.science/api/pith-number/F3WUDZEAMQX5PXM3BPC2ZJNQGX/graph.json","fetch_events":"https://pith.science/api/pith-number/F3WUDZEAMQX5PXM3BPC2ZJNQGX/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/F3WUDZEAMQX5PXM3BPC2ZJNQGX/action/timestamp_anchor","attest_storage":"https://pith.science/pith/F3WUDZEAMQX5PXM3BPC2ZJNQGX/action/storage_attestation","attest_author":"https://pith.science/pith/F3WUDZEAMQX5PXM3BPC2ZJNQGX/action/author_attestation","sign_citation":"https://pith.science/pith/F3WUDZEAMQX5PXM3BPC2ZJNQGX/action/citation_signature","submit_replication":"https://pith.science/pith/F3WUDZEAMQX5PXM3BPC2ZJNQGX/action/replication_record"}},"created_at":"2026-07-05T07:45:10.229804+00:00","updated_at":"2026-07-05T07:45:10.229804+00:00"}