{"bundle_type":"pith_open_graph_bundle","bundle_version":"1.0","pith_number":"pith:2025:SGZ57SD7DMJTLISNYY2HWNLQPH","short_pith_number":"pith:SGZ57SD7","canonical_record":{"source":{"id":"2502.20586","kind":"arxiv","version":3},"metadata":{"license":"http://creativecommons.org/licenses/by/4.0/","primary_cat":"cs.LG","submitted_at":"2025-02-27T23:01:31Z","cross_cats_sorted":[],"title_canon_sha256":"a0f5d050bb1fe9b5a2e2d8bf318427798fa48e8d52be52190c0fbfb4d109af70","abstract_canon_sha256":"37d11004505631b77cc250a6199ce6703c1a80c9b891ff52097619bdf6e43a78"},"schema_version":"1.0"},"canonical_sha256":"91b3dfc87f1b1335a24dc6347b357079f77f54ca9b98122bf15209e32267bbaa","source":{"kind":"arxiv","id":"2502.20586","version":3},"source_aliases":[{"alias_kind":"arxiv","alias_value":"2502.20586","created_at":"2026-07-05T11:59:57Z"},{"alias_kind":"arxiv_version","alias_value":"2502.20586v3","created_at":"2026-07-05T11:59:57Z"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2502.20586","created_at":"2026-07-05T11:59:57Z"},{"alias_kind":"pith_short_12","alias_value":"SGZ57SD7DMJT","created_at":"2026-07-05T11:59:57Z"},{"alias_kind":"pith_short_16","alias_value":"SGZ57SD7DMJTLISN","created_at":"2026-07-05T11:59:57Z"},{"alias_kind":"pith_short_8","alias_value":"SGZ57SD7","created_at":"2026-07-05T11:59:57Z"}],"events":[{"event_type":"record_created","subject_pith_number":"pith:2025:SGZ57SD7DMJTLISNYY2HWNLQPH","target":"record","payload":{"canonical_record":{"source":{"id":"2502.20586","kind":"arxiv","version":3},"metadata":{"license":"http://creativecommons.org/licenses/by/4.0/","primary_cat":"cs.LG","submitted_at":"2025-02-27T23:01:31Z","cross_cats_sorted":[],"title_canon_sha256":"a0f5d050bb1fe9b5a2e2d8bf318427798fa48e8d52be52190c0fbfb4d109af70","abstract_canon_sha256":"37d11004505631b77cc250a6199ce6703c1a80c9b891ff52097619bdf6e43a78"},"schema_version":"1.0"},"canonical_sha256":"91b3dfc87f1b1335a24dc6347b357079f77f54ca9b98122bf15209e32267bbaa","receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-07-05T11:59:57.960185Z","signature_b64":"MKE/Z8iqjBmEpNGWHmYRUkKB736yj7xk0D28h/4yAUldBaXNy69Ro0OLkW/iXtSwa3gWu1Hd29bUqhw67Q75Bg==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"91b3dfc87f1b1335a24dc6347b357079f77f54ca9b98122bf15209e32267bbaa","last_reissued_at":"2026-07-05T11:59:57.959721Z","signature_status":"signed_v1","first_computed_at":"2026-07-05T11:59:57.959721Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"source_kind":"arxiv","source_id":"2502.20586","source_version":3,"attestation_state":"computed"},"signer":{"signer_id":"pith.science","signer_type":"pith_registry","key_id":"pith-v1-2026-05","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"created_at":"2026-07-05T11:59:57Z","supersedes":[],"prev_event":null,"signature":{"signature_status":"signed_v1","algorithm":"ed25519","key_id":"pith-v1-2026-05","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54","signature_b64":"dBqqdf+gAkoIV/iecrWWvDb6epADxbhCUeo46KDaCYCHA4p0lJB2kQpw+FW5bVWEG2LTujrsjhX0g5KXUxewCA==","signed_message":"open_graph_event_sha256_bytes","signed_at":"2026-08-16T20:07:37.217524Z"},"content_sha256":"bb0199e1dcd83530789d2aac00b2f7c67e8d1df6624782792cb2a474cd944dfb","schema_version":"1.0","event_id":"sha256:bb0199e1dcd83530789d2aac00b2f7c67e8d1df6624782792cb2a474cd944dfb"},{"event_type":"graph_snapshot","subject_pith_number":"pith:2025:SGZ57SD7DMJTLISNYY2HWNLQPH","target":"graph","payload":{"graph_snapshot":{"paper":{"title":"Training LLMs with MXFP4","license":"http://creativecommons.org/licenses/by/4.0/","headline":"","cross_cats":[],"primary_cat":"cs.LG","authors_text":"Albert Tseng, Tao Yu, Youngsuk Park","submitted_at":"2025-02-27T23:01:31Z","abstract_excerpt":"Low precision (LP) datatypes such as MXFP4 can accelerate matrix multiplications (GEMMs) and reduce training costs. However, directly using MXFP4 instead of BF16 during training significantly degrades model quality. In this work, we present the first near-lossless training recipe that uses MXFP4 GEMMs, which are $2\\times$ faster than FP8 on supported hardware. Our key insight is to compute unbiased gradient estimates with stochastic rounding (SR), resulting in more accurate model updates. However, directly applying SR to MXFP4 can result in high variance from block-level outliers, harming conv"},"claims":{"count":0,"items":[],"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"source":{"id":"2502.20586","kind":"arxiv","version":3},"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/2502.20586/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"},"verdict_id":null},"signer":{"signer_id":"pith.science","signer_type":"pith_registry","key_id":"pith-v1-2026-05","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"created_at":"2026-07-05T11:59:57Z","supersedes":[],"prev_event":null,"signature":{"signature_status":"signed_v1","algorithm":"ed25519","key_id":"pith-v1-2026-05","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54","signature_b64":"WHFg9UdKsFFJcwnQeFqZ9CdfswPGDdFSobfy6JZtcVL/EKuqRIRp4oIBQhKOMhe9vQg6++chCN3rJ6nfx/enCA==","signed_message":"open_graph_event_sha256_bytes","signed_at":"2026-08-16T20:07:37.218327Z"},"content_sha256":"2c29efb01b633d3abcb141fafb6467a44da7efc3ce15da42d67140d48d143bb5","schema_version":"1.0","event_id":"sha256:2c29efb01b633d3abcb141fafb6467a44da7efc3ce15da42d67140d48d143bb5"}],"timestamp_proofs":[],"mirror_hints":[{"mirror_type":"https","name":"Pith Resolver","base_url":"https://pith.science","bundle_url":"https://pith.science/pith/SGZ57SD7DMJTLISNYY2HWNLQPH/bundle.json","state_url":"https://pith.science/pith/SGZ57SD7DMJTLISNYY2HWNLQPH/state.json","well_known_bundle_url":"https://pith.science/.well-known/pith/SGZ57SD7DMJTLISNYY2HWNLQPH/bundle.json","status":"primary"}],"public_keys":[{"key_id":"pith-v1-2026-05","algorithm":"ed25519","format":"raw","public_key_b64":"stVStoiQhXFxp4s2pdzPNoqVNBMojDU/fJ2db5S3CbM=","public_key_hex":"b2d552b68890857171a78b36a5dccf368a953413288c353f7c9d9d6f94b709b3","fingerprint_sha256_b32_first128bits":"RVFV5Z2OI2J3ZUO7ERDEBCYNKS","fingerprint_sha256_hex":"8d4b5ee74e4693bcd1df2446408b0d54","rotates_at":null,"url":"https://pith.science/pith-signing-key.json","notes":"Pith uses this Ed25519 key to sign canonical record SHA-256 digests. Verify with: ed25519_verify(public_key, message=canonical_sha256_bytes, signature=base64decode(signature_b64))."}],"merge_version":"pith-open-graph-merge-v1","built_at":"2026-08-16T20:07:37Z","links":{"resolver":"https://pith.science/pith/SGZ57SD7DMJTLISNYY2HWNLQPH","bundle":"https://pith.science/pith/SGZ57SD7DMJTLISNYY2HWNLQPH/bundle.json","state":"https://pith.science/pith/SGZ57SD7DMJTLISNYY2HWNLQPH/state.json","well_known_bundle":"https://pith.science/.well-known/pith/SGZ57SD7DMJTLISNYY2HWNLQPH/bundle.json"},"state":{"state_type":"pith_open_graph_state","state_version":"1.0","pith_number":"pith:2025:SGZ57SD7DMJTLISNYY2HWNLQPH","merge_version":"pith-open-graph-merge-v1","event_count":2,"valid_event_count":2,"invalid_event_count":0,"equivocation_count":0,"current":{"canonical_record":{"metadata":{"abstract_canon_sha256":"37d11004505631b77cc250a6199ce6703c1a80c9b891ff52097619bdf6e43a78","cross_cats_sorted":[],"license":"http://creativecommons.org/licenses/by/4.0/","primary_cat":"cs.LG","submitted_at":"2025-02-27T23:01:31Z","title_canon_sha256":"a0f5d050bb1fe9b5a2e2d8bf318427798fa48e8d52be52190c0fbfb4d109af70"},"schema_version":"1.0","source":{"id":"2502.20586","kind":"arxiv","version":3}},"source_aliases":[{"alias_kind":"arxiv","alias_value":"2502.20586","created_at":"2026-07-05T11:59:57Z"},{"alias_kind":"arxiv_version","alias_value":"2502.20586v3","created_at":"2026-07-05T11:59:57Z"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2502.20586","created_at":"2026-07-05T11:59:57Z"},{"alias_kind":"pith_short_12","alias_value":"SGZ57SD7DMJT","created_at":"2026-07-05T11:59:57Z"},{"alias_kind":"pith_short_16","alias_value":"SGZ57SD7DMJTLISN","created_at":"2026-07-05T11:59:57Z"},{"alias_kind":"pith_short_8","alias_value":"SGZ57SD7","created_at":"2026-07-05T11:59:57Z"}],"graph_snapshots":[{"event_id":"sha256:2c29efb01b633d3abcb141fafb6467a44da7efc3ce15da42d67140d48d143bb5","target":"graph","created_at":"2026-07-05T11:59:57Z","signer":{"key_id":"pith-v1-2026-05","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54","signer_id":"pith.science","signer_type":"pith_registry"},"payload":{"graph_snapshot":{"author_claims":{"count":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57","strong_count":0},"builder_version":"pith-number-builder-2026-05-17-v1","claims":{"count":0,"items":[],"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"formal_canon":{"evidence_count":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"integrity":{"available":true,"clean":true,"detectors_run":[],"endpoint":"/pith/2502.20586/integrity.json","findings":[],"snapshot_sha256":"c28c3603d3b5d939e8dc4c7e95fa8dfce3d595e45f758748cecf8e644a296938","summary":{"advisory":0,"by_detector":{},"critical":0,"informational":0}},"paper":{"abstract_excerpt":"Low precision (LP) datatypes such as MXFP4 can accelerate matrix multiplications (GEMMs) and reduce training costs. However, directly using MXFP4 instead of BF16 during training significantly degrades model quality. In this work, we present the first near-lossless training recipe that uses MXFP4 GEMMs, which are $2\\times$ faster than FP8 on supported hardware. Our key insight is to compute unbiased gradient estimates with stochastic rounding (SR), resulting in more accurate model updates. However, directly applying SR to MXFP4 can result in high variance from block-level outliers, harming conv","authors_text":"Albert Tseng, Tao Yu, Youngsuk Park","cross_cats":[],"headline":"","license":"http://creativecommons.org/licenses/by/4.0/","primary_cat":"cs.LG","submitted_at":"2025-02-27T23:01:31Z","title":"Training LLMs with MXFP4"},"references":{"count":0,"internal_anchors":0,"resolved_work":0,"sample":[],"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"source":{"id":"2502.20586","kind":"arxiv","version":3},"verdict":{"created_at":null,"id":null,"model_set":{},"one_line_summary":"","pipeline_version":null,"pith_extraction_headline":"","strongest_claim":"","weakest_assumption":""}},"verdict_id":null}}],"author_attestations":[],"timestamp_anchors":[],"storage_attestations":[],"citation_signatures":[],"replication_records":[],"corrections":[],"mirror_hints":[],"record_created":{"event_id":"sha256:bb0199e1dcd83530789d2aac00b2f7c67e8d1df6624782792cb2a474cd944dfb","target":"record","created_at":"2026-07-05T11:59:57Z","signer":{"key_id":"pith-v1-2026-05","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54","signer_id":"pith.science","signer_type":"pith_registry"},"payload":{"attestation_state":"computed","canonical_record":{"metadata":{"abstract_canon_sha256":"37d11004505631b77cc250a6199ce6703c1a80c9b891ff52097619bdf6e43a78","cross_cats_sorted":[],"license":"http://creativecommons.org/licenses/by/4.0/","primary_cat":"cs.LG","submitted_at":"2025-02-27T23:01:31Z","title_canon_sha256":"a0f5d050bb1fe9b5a2e2d8bf318427798fa48e8d52be52190c0fbfb4d109af70"},"schema_version":"1.0","source":{"id":"2502.20586","kind":"arxiv","version":3}},"canonical_sha256":"91b3dfc87f1b1335a24dc6347b357079f77f54ca9b98122bf15209e32267bbaa","receipt":{"algorithm":"ed25519","builder_version":"pith-number-builder-2026-05-17-v1","canonical_sha256":"91b3dfc87f1b1335a24dc6347b357079f77f54ca9b98122bf15209e32267bbaa","first_computed_at":"2026-07-05T11:59:57.959721Z","key_id":"pith-v1-2026-05","kind":"pith_receipt","last_reissued_at":"2026-07-05T11:59:57.959721Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54","receipt_version":"0.3","signature_b64":"MKE/Z8iqjBmEpNGWHmYRUkKB736yj7xk0D28h/4yAUldBaXNy69Ro0OLkW/iXtSwa3gWu1Hd29bUqhw67Q75Bg==","signature_status":"signed_v1","signed_at":"2026-07-05T11:59:57.960185Z","signed_message":"canonical_sha256_bytes"},"source_id":"2502.20586","source_kind":"arxiv","source_version":3}}},"equivocations":[],"invalid_events":[],"applied_event_ids":["sha256:bb0199e1dcd83530789d2aac00b2f7c67e8d1df6624782792cb2a474cd944dfb","sha256:2c29efb01b633d3abcb141fafb6467a44da7efc3ce15da42d67140d48d143bb5"],"state_sha256":"e99238f866e37fe674b297227bc96aac01c7162c1fdcb73b3a69b07526a198fb"},"bundle_signature":{"signature_status":"signed_v1","algorithm":"ed25519","key_id":"pith-v1-2026-05","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54","signature_b64":"GkwUXyQRjwXwMYv4CDx0AcuESM0gnYikNSXHxiySjPrhzsm5/oVmUd4jWqmuxUMlGks1qN6k1KvN857llG2RCg==","signed_message":"bundle_sha256_bytes","signed_at":"2026-08-16T20:07:37.222725Z","bundle_sha256":"3945d175be253efebb8e4e8e84bdfe981eaa6f782a72e9bfe9bca9cf38592492"}}