{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2026:4CRAALYV43ILG3VBJ7MKP2EKAV","short_pith_number":"pith:4CRAALYV","schema_version":"1.0","canonical_sha256":"e0a2002f15e6d0b36ea14fd8a7e88a057cb26629b23a0b06e6c8c4364077632a","source":{"kind":"arxiv","id":"2608.02091","version":1},"attestation_state":"computed","paper":{"title":"One QK Channel, Many Sources: Guarding Low-Precision Attention Collapse","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","headline":"","cross_cats":[],"primary_cat":"cs.LG","authors_text":"Dezhi Ran, Shuxiao Xie, Shuyang Xie, Tao Xie, Wei Yang, Yuan Cao","submitted_at":"2026-08-03T11:50:06Z","abstract_excerpt":"A bfloat16 transformer can train normally for many steps and then collapse abruptly. Distinct low-precision errors can trigger the same failure, leaving unclear whether each source needs its own repair or one shared route can be blocked. We isolate a reproduced GPT-2-class collapse to the streaming-softmax accumulator, where fp32 accumulation repairs it, and use the fault as an assay for moving controlled errors across sources. Errors placed outside attention still drive the same query-key (QK) spectral runaway, while correcting only QK keeps training stable with the source fault active. This "},"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":"2608.02091","kind":"arxiv","version":1},"metadata":{"license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","primary_cat":"cs.LG","submitted_at":"2026-08-03T11:50:06Z","cross_cats_sorted":[],"title_canon_sha256":"dbf3e1a491c783a8f28fb54a8a5e2393f2aff38e24f6f879401613577ce9ae9a","abstract_canon_sha256":"7aef3519790a5870ffe939e0f99807d40803e10f154eaa4c49082270dc509c38"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-08-04T02:10:49.800492Z","signature_b64":"zQ9PIVnts6VJ0zyFQauNmZ+t/2j90KIaKpRf8y/xTsO3Ndakr6NXBb8lYcWwL4kGJP8mbJb0Smy4UjXD97KPBw==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"e0a2002f15e6d0b36ea14fd8a7e88a057cb26629b23a0b06e6c8c4364077632a","last_reissued_at":"2026-08-04T02:10:49.798841Z","signature_status":"signed_v1","first_computed_at":"2026-08-04T02:10:49.798841Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"One QK Channel, Many Sources: Guarding Low-Precision Attention Collapse","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","headline":"","cross_cats":[],"primary_cat":"cs.LG","authors_text":"Dezhi Ran, Shuxiao Xie, Shuyang Xie, Tao Xie, Wei Yang, Yuan Cao","submitted_at":"2026-08-03T11:50:06Z","abstract_excerpt":"A bfloat16 transformer can train normally for many steps and then collapse abruptly. Distinct low-precision errors can trigger the same failure, leaving unclear whether each source needs its own repair or one shared route can be blocked. We isolate a reproduced GPT-2-class collapse to the streaming-softmax accumulator, where fp32 accumulation repairs it, and use the fault as an assay for moving controlled errors across sources. Errors placed outside attention still drive the same query-key (QK) spectral runaway, while correcting only QK keeps training stable with the source fault active. This "},"claims":{"count":0,"items":[],"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"source":{"id":"2608.02091","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/2608.02091/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":"2608.02091","created_at":"2026-08-04T02:10:49.800458+00:00"},{"alias_kind":"arxiv_version","alias_value":"2608.02091v1","created_at":"2026-08-04T02:10:49.800458+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2608.02091","created_at":"2026-08-04T02:10:49.800458+00:00"},{"alias_kind":"pith_short_12","alias_value":"4CRAALYV43IL","created_at":"2026-08-04T02:10:49.800458+00:00"},{"alias_kind":"pith_short_16","alias_value":"4CRAALYV43ILG3VB","created_at":"2026-08-04T02:10:49.800458+00:00"},{"alias_kind":"pith_short_8","alias_value":"4CRAALYV","created_at":"2026-08-04T02:10:49.800458+00:00"}],"events":[],"event_summary":{},"paper_claims":[],"inbound_citations":{"count":0,"internal_anchor_count":0,"sample":[]},"formal_canon":{"evidence_count":0,"sample":[],"anchors":[]},"links":{"html":"https://pith.science/pith/4CRAALYV43ILG3VBJ7MKP2EKAV","json":"https://pith.science/pith/4CRAALYV43ILG3VBJ7MKP2EKAV.json","graph_json":"https://pith.science/api/pith-number/4CRAALYV43ILG3VBJ7MKP2EKAV/graph.json","events_json":"https://pith.science/api/pith-number/4CRAALYV43ILG3VBJ7MKP2EKAV/events.json","paper":"https://pith.science/paper/4CRAALYV"},"agent_actions":{"view_html":"https://pith.science/pith/4CRAALYV43ILG3VBJ7MKP2EKAV","download_json":"https://pith.science/pith/4CRAALYV43ILG3VBJ7MKP2EKAV.json","view_paper":"https://pith.science/paper/4CRAALYV","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2608.02091&json=true","fetch_graph":"https://pith.science/api/pith-number/4CRAALYV43ILG3VBJ7MKP2EKAV/graph.json","fetch_events":"https://pith.science/api/pith-number/4CRAALYV43ILG3VBJ7MKP2EKAV/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/4CRAALYV43ILG3VBJ7MKP2EKAV/action/timestamp_anchor","attest_storage":"https://pith.science/pith/4CRAALYV43ILG3VBJ7MKP2EKAV/action/storage_attestation","attest_author":"https://pith.science/pith/4CRAALYV43ILG3VBJ7MKP2EKAV/action/author_attestation","sign_citation":"https://pith.science/pith/4CRAALYV43ILG3VBJ7MKP2EKAV/action/citation_signature","submit_replication":"https://pith.science/pith/4CRAALYV43ILG3VBJ7MKP2EKAV/action/replication_record"}},"created_at":"2026-08-04T02:10:49.800458+00:00","updated_at":"2026-08-04T02:10:49.800458+00:00"}