{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2025:3J46E52JFS4EUNMNP4S4UTJKUD","short_pith_number":"pith:3J46E52J","schema_version":"1.0","canonical_sha256":"da79e277492cb84a358d7f25ca4d2aa0e9fdb3cd1e29b406ecc824970d362f41","source":{"kind":"arxiv","id":"2504.18415","version":2},"attestation_state":"computed","paper":{"title":"BitNet v2: Native 4-bit Activations with Hadamard Transformation for 1-bit LLMs","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","headline":"","cross_cats":["cs.LG"],"primary_cat":"cs.CL","authors_text":"Furu Wei, Hongyu Wang, Shuming Ma","submitted_at":"2025-04-25T15:17:52Z","abstract_excerpt":"Efficient deployment of 1-bit Large Language Models (LLMs) is hindered by activation outliers, which complicate quantization to low bit-widths. We introduce BitNet v2, a novel framework enabling native 4-bit activation quantization for 1-bit LLMs. To tackle outliers in attention and feed-forward network activations, we propose H-BitLinear, a module applying an online Hadamard transformation prior to activation quantization. This transformation smooths sharp activation distributions into more Gaussian-like forms, suitable for low-bit representation. Experiments show BitNet v2 trained from scrat"},"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":"2504.18415","kind":"arxiv","version":2},"metadata":{"license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","primary_cat":"cs.CL","submitted_at":"2025-04-25T15:17:52Z","cross_cats_sorted":["cs.LG"],"title_canon_sha256":"3eef9237b01cde9725b007887dd090f2e7e4449fe66873d4d7bda240f7b4bbc3","abstract_canon_sha256":"af0d87215c3ac467daa754823025379b41717556beccc65c66f5368f4a444aa9"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-07-05T11:20:53.212965Z","signature_b64":"2eDIW9R1HxgOg6f+zol3fgg9s2B8KaVbzc5BSxMy1P3Enw2Y9c1fk8nsAuIrWLO+rJLn7r55kSG2AJ7JV8OMDg==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"da79e277492cb84a358d7f25ca4d2aa0e9fdb3cd1e29b406ecc824970d362f41","last_reissued_at":"2026-07-05T11:20:53.212452Z","signature_status":"signed_v1","first_computed_at":"2026-07-05T11:20:53.212452Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"BitNet v2: Native 4-bit Activations with Hadamard Transformation for 1-bit LLMs","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","headline":"","cross_cats":["cs.LG"],"primary_cat":"cs.CL","authors_text":"Furu Wei, Hongyu Wang, Shuming Ma","submitted_at":"2025-04-25T15:17:52Z","abstract_excerpt":"Efficient deployment of 1-bit Large Language Models (LLMs) is hindered by activation outliers, which complicate quantization to low bit-widths. We introduce BitNet v2, a novel framework enabling native 4-bit activation quantization for 1-bit LLMs. To tackle outliers in attention and feed-forward network activations, we propose H-BitLinear, a module applying an online Hadamard transformation prior to activation quantization. This transformation smooths sharp activation distributions into more Gaussian-like forms, suitable for low-bit representation. Experiments show BitNet v2 trained from scrat"},"claims":{"count":0,"items":[],"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"source":{"id":"2504.18415","kind":"arxiv","version":2},"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/2504.18415/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":"2504.18415","created_at":"2026-07-05T11:20:53.212520+00:00"},{"alias_kind":"arxiv_version","alias_value":"2504.18415v2","created_at":"2026-07-05T11:20:53.212520+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2504.18415","created_at":"2026-07-05T11:20:53.212520+00:00"},{"alias_kind":"pith_short_12","alias_value":"3J46E52JFS4E","created_at":"2026-07-05T11:20:53.212520+00:00"},{"alias_kind":"pith_short_16","alias_value":"3J46E52JFS4EUNMN","created_at":"2026-07-05T11:20:53.212520+00:00"},{"alias_kind":"pith_short_8","alias_value":"3J46E52J","created_at":"2026-07-05T11:20:53.212520+00:00"}],"events":[],"event_summary":{},"paper_claims":[],"inbound_citations":{"count":6,"internal_anchor_count":0,"sample":[{"citing_arxiv_id":"2606.26650","citing_title":"CAT-Q: Cost-efficient and Accurate Ternary Quantization for LLMs","ref_index":19,"is_internal_anchor":false},{"citing_arxiv_id":"2606.13054","citing_title":"TWLA: Achieving Ternary Weights and Low-Bit Activations for LLMs via Post-Training Quantization","ref_index":21,"is_internal_anchor":false},{"citing_arxiv_id":"2606.10520","citing_title":"UniSVQ: 2-bit Unified Scalar-Vector Quantization","ref_index":10,"is_internal_anchor":false},{"citing_arxiv_id":"2606.10520","citing_title":"UniSVQ: 2-bit Unified Scalar-Vector Quantization","ref_index":10,"is_internal_anchor":false},{"citing_arxiv_id":"2509.23582","citing_title":"RobuQ: Pushing DiTs to W1.58A2 via Robust Activation Quantization","ref_index":58,"is_internal_anchor":false},{"citing_arxiv_id":"2604.19167","citing_title":"LBLLM: Lightweight Binarization of Large Language Models via Three-Stage Distillation","ref_index":91,"is_internal_anchor":false}]},"formal_canon":{"evidence_count":0,"sample":[],"anchors":[]},"links":{"html":"https://pith.science/pith/3J46E52JFS4EUNMNP4S4UTJKUD","json":"https://pith.science/pith/3J46E52JFS4EUNMNP4S4UTJKUD.json","graph_json":"https://pith.science/api/pith-number/3J46E52JFS4EUNMNP4S4UTJKUD/graph.json","events_json":"https://pith.science/api/pith-number/3J46E52JFS4EUNMNP4S4UTJKUD/events.json","paper":"https://pith.science/paper/3J46E52J"},"agent_actions":{"view_html":"https://pith.science/pith/3J46E52JFS4EUNMNP4S4UTJKUD","download_json":"https://pith.science/pith/3J46E52JFS4EUNMNP4S4UTJKUD.json","view_paper":"https://pith.science/paper/3J46E52J","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2504.18415&json=true","fetch_graph":"https://pith.science/api/pith-number/3J46E52JFS4EUNMNP4S4UTJKUD/graph.json","fetch_events":"https://pith.science/api/pith-number/3J46E52JFS4EUNMNP4S4UTJKUD/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/3J46E52JFS4EUNMNP4S4UTJKUD/action/timestamp_anchor","attest_storage":"https://pith.science/pith/3J46E52JFS4EUNMNP4S4UTJKUD/action/storage_attestation","attest_author":"https://pith.science/pith/3J46E52JFS4EUNMNP4S4UTJKUD/action/author_attestation","sign_citation":"https://pith.science/pith/3J46E52JFS4EUNMNP4S4UTJKUD/action/citation_signature","submit_replication":"https://pith.science/pith/3J46E52JFS4EUNMNP4S4UTJKUD/action/replication_record"}},"created_at":"2026-07-05T11:20:53.212520+00:00","updated_at":"2026-07-05T11:20:53.212520+00:00"}