{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2026:3R3RAISTDKFJS36FY7L4YL3IVJ","short_pith_number":"pith:3R3RAIST","schema_version":"1.0","canonical_sha256":"dc771022531a8a996fc5c7d7cc2f68aa7767c624f0fc5dafe2389f1550005941","source":{"kind":"arxiv","id":"2606.00539","version":1},"attestation_state":"computed","paper":{"title":"GNMR: Runtime Stability Control for Low-Precision Large Language Model Training","license":"http://creativecommons.org/licenses/by/4.0/","headline":"","cross_cats":["math.OC","stat.ML"],"primary_cat":"cs.LG","authors_text":"Boao Kong, Engao Zhang, Guohong Li, Kun Yuan, Weichen Jia, Yao Wang, Yaoyuan Wang, Yonghan Dong, Yunke Peng","submitted_at":"2026-05-30T05:11:13Z","abstract_excerpt":"Training stability is a key bottleneck in low-precision language model training: efficient low-cost paths can still produce short-lived numerical risks at a small set of operators. We formulate this as runtime stability control and present Gradient Norm-to-Mean Ratio (GNMR), a lightweight controller that compares each recoverable unit's current gradient norm with its historical mean. Together with $\\Delta$-GNMR for abrupt short-window increases, GNMR maps local risk signals to bounded recovery actions under a hard $\\mathrm{maxO}$ budget and a short lock interval, without changing the numerical"},"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":"2606.00539","kind":"arxiv","version":1},"metadata":{"license":"http://creativecommons.org/licenses/by/4.0/","primary_cat":"cs.LG","submitted_at":"2026-05-30T05:11:13Z","cross_cats_sorted":["math.OC","stat.ML"],"title_canon_sha256":"8004a7a4102d496e89fd8b3701d515120121a871ce920e3947ba9f312f55beac","abstract_canon_sha256":"0401f827309c542b473c6e5746e57e61bcd0b5a29b266c6b0594255d825152f4"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-06-02T01:03:57.604546Z","signature_b64":"ZAgUxAq+lqSMvEhzPHXzvlseIbT/EUxhN4XRjV3C/MKTWh9M4uY8p5iZNvBkg4DNRcCREeOuvCx0ZwDDEeUMDQ==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"dc771022531a8a996fc5c7d7cc2f68aa7767c624f0fc5dafe2389f1550005941","last_reissued_at":"2026-06-02T01:03:57.604143Z","signature_status":"signed_v1","first_computed_at":"2026-06-02T01:03:57.604143Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"GNMR: Runtime Stability Control for Low-Precision Large Language Model Training","license":"http://creativecommons.org/licenses/by/4.0/","headline":"","cross_cats":["math.OC","stat.ML"],"primary_cat":"cs.LG","authors_text":"Boao Kong, Engao Zhang, Guohong Li, Kun Yuan, Weichen Jia, Yao Wang, Yaoyuan Wang, Yonghan Dong, Yunke Peng","submitted_at":"2026-05-30T05:11:13Z","abstract_excerpt":"Training stability is a key bottleneck in low-precision language model training: efficient low-cost paths can still produce short-lived numerical risks at a small set of operators. We formulate this as runtime stability control and present Gradient Norm-to-Mean Ratio (GNMR), a lightweight controller that compares each recoverable unit's current gradient norm with its historical mean. Together with $\\Delta$-GNMR for abrupt short-window increases, GNMR maps local risk signals to bounded recovery actions under a hard $\\mathrm{maxO}$ budget and a short lock interval, without changing the numerical"},"claims":{"count":0,"items":[],"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"source":{"id":"2606.00539","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/2606.00539/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":"2606.00539","created_at":"2026-06-02T01:03:57.604198+00:00"},{"alias_kind":"arxiv_version","alias_value":"2606.00539v1","created_at":"2026-06-02T01:03:57.604198+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2606.00539","created_at":"2026-06-02T01:03:57.604198+00:00"},{"alias_kind":"pith_short_12","alias_value":"3R3RAISTDKFJ","created_at":"2026-06-02T01:03:57.604198+00:00"},{"alias_kind":"pith_short_16","alias_value":"3R3RAISTDKFJS36F","created_at":"2026-06-02T01:03:57.604198+00:00"},{"alias_kind":"pith_short_8","alias_value":"3R3RAIST","created_at":"2026-06-02T01:03:57.604198+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/3R3RAISTDKFJS36FY7L4YL3IVJ","json":"https://pith.science/pith/3R3RAISTDKFJS36FY7L4YL3IVJ.json","graph_json":"https://pith.science/api/pith-number/3R3RAISTDKFJS36FY7L4YL3IVJ/graph.json","events_json":"https://pith.science/api/pith-number/3R3RAISTDKFJS36FY7L4YL3IVJ/events.json","paper":"https://pith.science/paper/3R3RAIST"},"agent_actions":{"view_html":"https://pith.science/pith/3R3RAISTDKFJS36FY7L4YL3IVJ","download_json":"https://pith.science/pith/3R3RAISTDKFJS36FY7L4YL3IVJ.json","view_paper":"https://pith.science/paper/3R3RAIST","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2606.00539&json=true","fetch_graph":"https://pith.science/api/pith-number/3R3RAISTDKFJS36FY7L4YL3IVJ/graph.json","fetch_events":"https://pith.science/api/pith-number/3R3RAISTDKFJS36FY7L4YL3IVJ/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/3R3RAISTDKFJS36FY7L4YL3IVJ/action/timestamp_anchor","attest_storage":"https://pith.science/pith/3R3RAISTDKFJS36FY7L4YL3IVJ/action/storage_attestation","attest_author":"https://pith.science/pith/3R3RAISTDKFJS36FY7L4YL3IVJ/action/author_attestation","sign_citation":"https://pith.science/pith/3R3RAISTDKFJS36FY7L4YL3IVJ/action/citation_signature","submit_replication":"https://pith.science/pith/3R3RAISTDKFJS36FY7L4YL3IVJ/action/replication_record"}},"created_at":"2026-06-02T01:03:57.604198+00:00","updated_at":"2026-06-02T01:03:57.604198+00:00"}