{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2024:AFOMRFHZZBTAVDFWFPPOLMMZTV","short_pith_number":"pith:AFOMRFHZ","schema_version":"1.0","canonical_sha256":"015cc894f9c8660a8cb62bdee5b1999d76ce4bcfcbccf2fc6a8c7da5556cd6dd","source":{"kind":"arxiv","id":"2404.05405","version":1},"attestation_state":"computed","paper":{"title":"Physics of Language Models: Part 3.3, Knowledge Capacity Scaling Laws","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","headline":"","cross_cats":["cs.AI","cs.LG"],"primary_cat":"cs.CL","authors_text":"Yuanzhi Li, Zeyuan Allen-Zhu","submitted_at":"2024-04-08T11:11:31Z","abstract_excerpt":"Scaling laws describe the relationship between the size of language models and their capabilities. Unlike prior studies that evaluate a model's capability via loss or benchmarks, we estimate the number of knowledge bits a model stores. We focus on factual knowledge represented as tuples, such as (USA, capital, Washington D.C.) from a Wikipedia page. Through multiple controlled datasets, we establish that language models can and only can store 2 bits of knowledge per parameter, even when quantized to int8, and such knowledge can be flexibly extracted for downstream applications. Consequently, a"},"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":"2404.05405","kind":"arxiv","version":1},"metadata":{"license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","primary_cat":"cs.CL","submitted_at":"2024-04-08T11:11:31Z","cross_cats_sorted":["cs.AI","cs.LG"],"title_canon_sha256":"e6036c4b12d24a2ca56b17c43508a76d8ca2a7f0c6e1e52ea315bba845e1f2ed","abstract_canon_sha256":"d3275bbe306ab87fbd491098c0390ccab7922604d3a8f4b51b988cc348a67455"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-07-05T08:05:30.283307Z","signature_b64":"JLVBgtseKKKyecoZU/oyomLnk4TBbmQSEYB0H5wrLeBJRKJ+Ll/8E5UqQrClqooD6HX4K2EJnytKrSFAwaQTBQ==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"015cc894f9c8660a8cb62bdee5b1999d76ce4bcfcbccf2fc6a8c7da5556cd6dd","last_reissued_at":"2026-07-05T08:05:30.282881Z","signature_status":"signed_v1","first_computed_at":"2026-07-05T08:05:30.282881Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"Physics of Language Models: Part 3.3, Knowledge Capacity Scaling Laws","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","headline":"","cross_cats":["cs.AI","cs.LG"],"primary_cat":"cs.CL","authors_text":"Yuanzhi Li, Zeyuan Allen-Zhu","submitted_at":"2024-04-08T11:11:31Z","abstract_excerpt":"Scaling laws describe the relationship between the size of language models and their capabilities. Unlike prior studies that evaluate a model's capability via loss or benchmarks, we estimate the number of knowledge bits a model stores. We focus on factual knowledge represented as tuples, such as (USA, capital, Washington D.C.) from a Wikipedia page. Through multiple controlled datasets, we establish that language models can and only can store 2 bits of knowledge per parameter, even when quantized to int8, and such knowledge can be flexibly extracted for downstream applications. Consequently, a"},"claims":{"count":0,"items":[],"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"source":{"id":"2404.05405","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/2404.05405/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":"2404.05405","created_at":"2026-07-05T08:05:30.282930+00:00"},{"alias_kind":"arxiv_version","alias_value":"2404.05405v1","created_at":"2026-07-05T08:05:30.282930+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2404.05405","created_at":"2026-07-05T08:05:30.282930+00:00"},{"alias_kind":"pith_short_12","alias_value":"AFOMRFHZZBTA","created_at":"2026-07-05T08:05:30.282930+00:00"},{"alias_kind":"pith_short_16","alias_value":"AFOMRFHZZBTAVDFW","created_at":"2026-07-05T08:05:30.282930+00:00"},{"alias_kind":"pith_short_8","alias_value":"AFOMRFHZ","created_at":"2026-07-05T08:05:30.282930+00:00"}],"events":[],"event_summary":{},"paper_claims":[],"inbound_citations":{"count":14,"internal_anchor_count":0,"sample":[{"citing_arxiv_id":"2606.19172","citing_title":"User as Engram: Internalizing Per-User Memory as Local Parametric Edits","ref_index":81,"is_internal_anchor":false},{"citing_arxiv_id":"2606.04662","citing_title":"Why Muon Outperforms Adam: A Curvature Perspective","ref_index":65,"is_internal_anchor":false},{"citing_arxiv_id":"2606.03142","citing_title":"Disentangling Visual and Factual Correctness in LVLMs' Visualization Literacy","ref_index":60,"is_internal_anchor":false},{"citing_arxiv_id":"2605.26097","citing_title":"Forgetting in Language Models: Capacity, Optimization, and Self-Generated Replay","ref_index":19,"is_internal_anchor":false},{"citing_arxiv_id":"2606.00570","citing_title":"Revisiting Parameter-Based Knowledge Editing in Large Language Models: Theoretical Limits and Empirical Evidence","ref_index":2,"is_internal_anchor":false},{"citing_arxiv_id":"2605.18732","citing_title":"Predictable Confabulations: Factual Recall by LLMs Scales with Model Size and Topic Frequency","ref_index":4,"is_internal_anchor":false},{"citing_arxiv_id":"2511.21613","citing_title":"Beyond URLs: Metadata Diversity and Position for Efficient LLM Pretraining","ref_index":3,"is_internal_anchor":false},{"citing_arxiv_id":"2603.26554","citing_title":"Sharp Capacity Scaling of Spectral Optimizers in Learning Associative Memory","ref_index":1,"is_internal_anchor":false},{"citing_arxiv_id":"2605.12426","citing_title":"Geometric Factual Recall in Transformers","ref_index":18,"is_internal_anchor":false},{"citing_arxiv_id":"2605.10129","citing_title":"Synthetic Pre-Pre-Training Improves Language Model Robustness to Noisy Pre-Training Data","ref_index":52,"is_internal_anchor":false},{"citing_arxiv_id":"2605.09471","citing_title":"The Statistical Cost of Adaptation in Multi-Source Transfer Learning","ref_index":149,"is_internal_anchor":false},{"citing_arxiv_id":"2605.09724","citing_title":"Model Capacity Determines Grokking through Competing Memorisation and Generalisation Speeds","ref_index":22,"is_internal_anchor":false},{"citing_arxiv_id":"2605.05189","citing_title":"Sharp Capacity Thresholds in Linear Associative Memory: From Winner-Take-All to Listwise Retrieval","ref_index":37,"is_internal_anchor":false},{"citing_arxiv_id":"2604.08519","citing_title":"Cram Less to Fit More: Training Data Pruning Improves Memorization of Facts","ref_index":2,"is_internal_anchor":false}]},"formal_canon":{"evidence_count":0,"sample":[],"anchors":[]},"links":{"html":"https://pith.science/pith/AFOMRFHZZBTAVDFWFPPOLMMZTV","json":"https://pith.science/pith/AFOMRFHZZBTAVDFWFPPOLMMZTV.json","graph_json":"https://pith.science/api/pith-number/AFOMRFHZZBTAVDFWFPPOLMMZTV/graph.json","events_json":"https://pith.science/api/pith-number/AFOMRFHZZBTAVDFWFPPOLMMZTV/events.json","paper":"https://pith.science/paper/AFOMRFHZ"},"agent_actions":{"view_html":"https://pith.science/pith/AFOMRFHZZBTAVDFWFPPOLMMZTV","download_json":"https://pith.science/pith/AFOMRFHZZBTAVDFWFPPOLMMZTV.json","view_paper":"https://pith.science/paper/AFOMRFHZ","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2404.05405&json=true","fetch_graph":"https://pith.science/api/pith-number/AFOMRFHZZBTAVDFWFPPOLMMZTV/graph.json","fetch_events":"https://pith.science/api/pith-number/AFOMRFHZZBTAVDFWFPPOLMMZTV/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/AFOMRFHZZBTAVDFWFPPOLMMZTV/action/timestamp_anchor","attest_storage":"https://pith.science/pith/AFOMRFHZZBTAVDFWFPPOLMMZTV/action/storage_attestation","attest_author":"https://pith.science/pith/AFOMRFHZZBTAVDFWFPPOLMMZTV/action/author_attestation","sign_citation":"https://pith.science/pith/AFOMRFHZZBTAVDFWFPPOLMMZTV/action/citation_signature","submit_replication":"https://pith.science/pith/AFOMRFHZZBTAVDFWFPPOLMMZTV/action/replication_record"}},"created_at":"2026-07-05T08:05:30.282930+00:00","updated_at":"2026-07-05T08:05:30.282930+00:00"}