{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2026:J3QPRZNSIGZBNFFYACBNDQS3TH","short_pith_number":"pith:J3QPRZNS","schema_version":"1.0","canonical_sha256":"4ee0f8e5b241b21694b80082d1c25b99d9a6fe461cff5924752904c3d8cbf853","source":{"kind":"arxiv","id":"2606.20381","version":1},"attestation_state":"computed","paper":{"title":"Rethinking Shrinkage Bias in LLM FP4 Pretraining: Geometric Origin, Systemic Impact, and UFP4 Recipe","license":"http://creativecommons.org/licenses/by/4.0/","headline":"","cross_cats":[],"primary_cat":"cs.AI","authors_text":"Changxin Tian, Chaofan Yu, Haitao Zhang, Jia Liu, Jun Zhou, Kunlong Chen, Mingliang Gong, Peijie Jiang, Qian Zhao, Zhiqiang Zhang, Zhonghui Jiang, Ziqi Liu","submitted_at":"2026-06-18T15:40:51Z","abstract_excerpt":"FP4 training promises substantial reductions in memory and computation cost for LLM pretraining, yet current FP4 hardware paths and recipes, including NVIDIA Blackwell/Rubin-class systems and AMD MI350-series GPUs, remain centered on E2M1 data elements. In this study, we identify a fundamental limitation of that choice: non-uniform formats such as E2M1 inherently suffer from Shrinkage Bias, a systematic negative rounding error caused by the geometric asymmetry of their representable bins. We show that this bias accumulates multiplicatively across layers and is amplified by the Random Hadamard "},"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":"2606.20381","kind":"arxiv","version":1},"metadata":{"license":"http://creativecommons.org/licenses/by/4.0/","primary_cat":"cs.AI","submitted_at":"2026-06-18T15:40:51Z","cross_cats_sorted":[],"title_canon_sha256":"191b801f45a19d35e451ff65d446780c75643c88a2c16a2f4b3175676a986426","abstract_canon_sha256":"014a66f4291de939e377f32b4350aa8aa05c5cb0bf1a09e3a57b9a477efbda6b"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-06-19T16:13:11.028364Z","signature_b64":"fYWoKVjNqJgIHwHVSoIcsldVFRalreP6Q4AZMCGkSbpSjsqJN5JNW2SZASW5F5Bw9opPCwADQW32jO7wUoU8Ag==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"4ee0f8e5b241b21694b80082d1c25b99d9a6fe461cff5924752904c3d8cbf853","last_reissued_at":"2026-06-19T16:13:11.027940Z","signature_status":"signed_v1","first_computed_at":"2026-06-19T16:13:11.027940Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"Rethinking Shrinkage Bias in LLM FP4 Pretraining: Geometric Origin, Systemic Impact, and UFP4 Recipe","license":"http://creativecommons.org/licenses/by/4.0/","headline":"","cross_cats":[],"primary_cat":"cs.AI","authors_text":"Changxin Tian, Chaofan Yu, Haitao Zhang, Jia Liu, Jun Zhou, Kunlong Chen, Mingliang Gong, Peijie Jiang, Qian Zhao, Zhiqiang Zhang, Zhonghui Jiang, Ziqi Liu","submitted_at":"2026-06-18T15:40:51Z","abstract_excerpt":"FP4 training promises substantial reductions in memory and computation cost for LLM pretraining, yet current FP4 hardware paths and recipes, including NVIDIA Blackwell/Rubin-class systems and AMD MI350-series GPUs, remain centered on E2M1 data elements. In this study, we identify a fundamental limitation of that choice: non-uniform formats such as E2M1 inherently suffer from Shrinkage Bias, a systematic negative rounding error caused by the geometric asymmetry of their representable bins. We show that this bias accumulates multiplicatively across layers and is amplified by the Random Hadamard "},"claims":{"count":0,"items":[],"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"source":{"id":"2606.20381","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/2606.20381/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":"2606.20381","created_at":"2026-06-19T16:13:11.027996+00:00"},{"alias_kind":"arxiv_version","alias_value":"2606.20381v1","created_at":"2026-06-19T16:13:11.027996+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2606.20381","created_at":"2026-06-19T16:13:11.027996+00:00"},{"alias_kind":"pith_short_12","alias_value":"J3QPRZNSIGZB","created_at":"2026-06-19T16:13:11.027996+00:00"},{"alias_kind":"pith_short_16","alias_value":"J3QPRZNSIGZBNFFY","created_at":"2026-06-19T16:13:11.027996+00:00"},{"alias_kind":"pith_short_8","alias_value":"J3QPRZNS","created_at":"2026-06-19T16:13:11.027996+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/J3QPRZNSIGZBNFFYACBNDQS3TH","json":"https://pith.science/pith/J3QPRZNSIGZBNFFYACBNDQS3TH.json","graph_json":"https://pith.science/api/pith-number/J3QPRZNSIGZBNFFYACBNDQS3TH/graph.json","events_json":"https://pith.science/api/pith-number/J3QPRZNSIGZBNFFYACBNDQS3TH/events.json","paper":"https://pith.science/paper/J3QPRZNS"},"agent_actions":{"view_html":"https://pith.science/pith/J3QPRZNSIGZBNFFYACBNDQS3TH","download_json":"https://pith.science/pith/J3QPRZNSIGZBNFFYACBNDQS3TH.json","view_paper":"https://pith.science/paper/J3QPRZNS","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2606.20381&json=true","fetch_graph":"https://pith.science/api/pith-number/J3QPRZNSIGZBNFFYACBNDQS3TH/graph.json","fetch_events":"https://pith.science/api/pith-number/J3QPRZNSIGZBNFFYACBNDQS3TH/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/J3QPRZNSIGZBNFFYACBNDQS3TH/action/timestamp_anchor","attest_storage":"https://pith.science/pith/J3QPRZNSIGZBNFFYACBNDQS3TH/action/storage_attestation","attest_author":"https://pith.science/pith/J3QPRZNSIGZBNFFYACBNDQS3TH/action/author_attestation","sign_citation":"https://pith.science/pith/J3QPRZNSIGZBNFFYACBNDQS3TH/action/citation_signature","submit_replication":"https://pith.science/pith/J3QPRZNSIGZBNFFYACBNDQS3TH/action/replication_record"}},"created_at":"2026-06-19T16:13:11.027996+00:00","updated_at":"2026-06-19T16:13:11.027996+00:00"}