{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2026:MOLE3OGU4HQGG6MW5UVRNKCSOU","short_pith_number":"pith:MOLE3OGU","schema_version":"1.0","canonical_sha256":"63964db8d4e1e0637996ed2b16a852752f1931d336848397f45ecf771ed99ec1","source":{"kind":"arxiv","id":"2605.26189","version":1},"attestation_state":"computed","paper":{"title":"Max-Window Scale Estimation for Near-Lossless HiF8 W8A8 Quantization-Aware Training","license":"http://creativecommons.org/licenses/by/4.0/","headline":"","cross_cats":["cs.AI"],"primary_cat":"cs.LG","authors_text":"Fan Zhang, Jie Sun, Jinquan Shi, Li Zhou, Yingying Cheng, Zhaoyi Sun, Zhiyang He","submitted_at":"2026-05-25T09:19:57Z","abstract_excerpt":"Quantization-aware training (QAT) with low-bit floating-point formats enables efficient LLM deployment, yet introduces subtle failure modes invisible to standard training metrics. We present a systematic study of HiF8 W8A8 QAT for OpenPangu-Embedded-1B through the lens of Delayed Tensor Scaling (DTS). Across eight controlled experiments, we identify and disentangle two orthogonal failure modes: (i)amax saturation, where delayed scale estimates silently corrupt knowledge-sensitive representations via forward-pass clipping, and (ii)catastrophic forgetting, where an aggressive learning rate overw"},"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":"2605.26189","kind":"arxiv","version":1},"metadata":{"license":"http://creativecommons.org/licenses/by/4.0/","primary_cat":"cs.LG","submitted_at":"2026-05-25T09:19:57Z","cross_cats_sorted":["cs.AI"],"title_canon_sha256":"c5c51bcf111057ffd793140b69d74011304e5f2b9d4cd32b26a204f7a4091618","abstract_canon_sha256":"1309183e0665f384fa2d44219c08f65299d7f1f861b28bc2ac95af88f0ac8110"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-05-27T00:04:28.020778Z","signature_b64":"xfsGlHuUNOtjjq6SqQR3HU5Xx+yopNFalA/74pqKlFL/xPMBKQ/S7DSPpflPhIxbD9EutiFOy++uhiquZzftBA==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"63964db8d4e1e0637996ed2b16a852752f1931d336848397f45ecf771ed99ec1","last_reissued_at":"2026-05-27T00:04:28.020229Z","signature_status":"signed_v1","first_computed_at":"2026-05-27T00:04:28.020229Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"Max-Window Scale Estimation for Near-Lossless HiF8 W8A8 Quantization-Aware Training","license":"http://creativecommons.org/licenses/by/4.0/","headline":"","cross_cats":["cs.AI"],"primary_cat":"cs.LG","authors_text":"Fan Zhang, Jie Sun, Jinquan Shi, Li Zhou, Yingying Cheng, Zhaoyi Sun, Zhiyang He","submitted_at":"2026-05-25T09:19:57Z","abstract_excerpt":"Quantization-aware training (QAT) with low-bit floating-point formats enables efficient LLM deployment, yet introduces subtle failure modes invisible to standard training metrics. We present a systematic study of HiF8 W8A8 QAT for OpenPangu-Embedded-1B through the lens of Delayed Tensor Scaling (DTS). Across eight controlled experiments, we identify and disentangle two orthogonal failure modes: (i)amax saturation, where delayed scale estimates silently corrupt knowledge-sensitive representations via forward-pass clipping, and (ii)catastrophic forgetting, where an aggressive learning rate overw"},"claims":{"count":0,"items":[],"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"source":{"id":"2605.26189","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/2605.26189/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":"2605.26189","created_at":"2026-05-27T00:04:28.020300+00:00"},{"alias_kind":"arxiv_version","alias_value":"2605.26189v1","created_at":"2026-05-27T00:04:28.020300+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2605.26189","created_at":"2026-05-27T00:04:28.020300+00:00"},{"alias_kind":"pith_short_12","alias_value":"MOLE3OGU4HQG","created_at":"2026-05-27T00:04:28.020300+00:00"},{"alias_kind":"pith_short_16","alias_value":"MOLE3OGU4HQGG6MW","created_at":"2026-05-27T00:04:28.020300+00:00"},{"alias_kind":"pith_short_8","alias_value":"MOLE3OGU","created_at":"2026-05-27T00:04:28.020300+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/MOLE3OGU4HQGG6MW5UVRNKCSOU","json":"https://pith.science/pith/MOLE3OGU4HQGG6MW5UVRNKCSOU.json","graph_json":"https://pith.science/api/pith-number/MOLE3OGU4HQGG6MW5UVRNKCSOU/graph.json","events_json":"https://pith.science/api/pith-number/MOLE3OGU4HQGG6MW5UVRNKCSOU/events.json","paper":"https://pith.science/paper/MOLE3OGU"},"agent_actions":{"view_html":"https://pith.science/pith/MOLE3OGU4HQGG6MW5UVRNKCSOU","download_json":"https://pith.science/pith/MOLE3OGU4HQGG6MW5UVRNKCSOU.json","view_paper":"https://pith.science/paper/MOLE3OGU","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2605.26189&json=true","fetch_graph":"https://pith.science/api/pith-number/MOLE3OGU4HQGG6MW5UVRNKCSOU/graph.json","fetch_events":"https://pith.science/api/pith-number/MOLE3OGU4HQGG6MW5UVRNKCSOU/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/MOLE3OGU4HQGG6MW5UVRNKCSOU/action/timestamp_anchor","attest_storage":"https://pith.science/pith/MOLE3OGU4HQGG6MW5UVRNKCSOU/action/storage_attestation","attest_author":"https://pith.science/pith/MOLE3OGU4HQGG6MW5UVRNKCSOU/action/author_attestation","sign_citation":"https://pith.science/pith/MOLE3OGU4HQGG6MW5UVRNKCSOU/action/citation_signature","submit_replication":"https://pith.science/pith/MOLE3OGU4HQGG6MW5UVRNKCSOU/action/replication_record"}},"created_at":"2026-05-27T00:04:28.020300+00:00","updated_at":"2026-05-27T00:04:28.020300+00:00"}