{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2023:XTLBBB6Y4KTUBXU5EM27RGCOF7","short_pith_number":"pith:XTLBBB6Y","schema_version":"1.0","canonical_sha256":"bcd61087d8e2a740de9d2335f8984e2ffad77aecf94016c81c774f0aefaddc2e","source":{"kind":"arxiv","id":"2309.10668","version":2},"attestation_state":"computed","paper":{"title":"Language Modeling Is Compression","license":"http://creativecommons.org/licenses/by/4.0/","headline":"Large language models trained on text compress images and audio better than specialized tools.","cross_cats":["cs.AI","cs.CL","cs.IT","math.IT"],"primary_cat":"cs.LG","authors_text":"Anian Ruoss, Christopher Mattern, Elliot Catt, Gr\\'egoire Del\\'etang, Joel Veness, Jordi Grau-Moya, Laurent Orseau, Li Kevin Wenliang, Marcus Hutter, Matthew Aitchison, Paul-Ambroise Duquenne, Tim Genewein","submitted_at":"2023-09-19T14:50:38Z","abstract_excerpt":"It has long been established that predictive models can be transformed into lossless compressors and vice versa. Incidentally, in recent years, the machine learning community has focused on training increasingly large and powerful self-supervised (language) models. Since these large language models exhibit impressive predictive capabilities, they are well-positioned to be strong compressors. In this work, we advocate for viewing the prediction problem through the lens of compression and evaluate the compression capabilities of large (foundation) models. We show that large language models are p"},"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":true,"formal_links_present":false},"canonical_record":{"source":{"id":"2309.10668","kind":"arxiv","version":2},"metadata":{"license":"http://creativecommons.org/licenses/by/4.0/","primary_cat":"cs.LG","submitted_at":"2023-09-19T14:50:38Z","cross_cats_sorted":["cs.AI","cs.CL","cs.IT","math.IT"],"title_canon_sha256":"6e120a3a5dcd5a2ea8b8e58a3af16ddbf5cf63cc0fa224a78c89c0a65669247e","abstract_canon_sha256":"def926c0e7ca4abc4365977dcb574ca4dabb545c6b3e14e3b6b81a9cc38c332a"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-05-17T23:38:12.795772Z","signature_b64":"QtCKI/0PiLLS6CBep3ynAe9GZsnKyAgf7rRCFdvzTKJYIgGVXl8LNEjNtvN9FLJhOSYGsX+fPW433Kt50fciDw==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"bcd61087d8e2a740de9d2335f8984e2ffad77aecf94016c81c774f0aefaddc2e","last_reissued_at":"2026-05-17T23:38:12.795179Z","signature_status":"signed_v1","first_computed_at":"2026-05-17T23:38:12.795179Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"Language Modeling Is Compression","license":"http://creativecommons.org/licenses/by/4.0/","headline":"Large language models trained on text compress images and audio better than specialized tools.","cross_cats":["cs.AI","cs.CL","cs.IT","math.IT"],"primary_cat":"cs.LG","authors_text":"Anian Ruoss, Christopher Mattern, Elliot Catt, Gr\\'egoire Del\\'etang, Joel Veness, Jordi Grau-Moya, Laurent Orseau, Li Kevin Wenliang, Marcus Hutter, Matthew Aitchison, Paul-Ambroise Duquenne, Tim Genewein","submitted_at":"2023-09-19T14:50:38Z","abstract_excerpt":"It has long been established that predictive models can be transformed into lossless compressors and vice versa. Incidentally, in recent years, the machine learning community has focused on training increasingly large and powerful self-supervised (language) models. Since these large language models exhibit impressive predictive capabilities, they are well-positioned to be strong compressors. In this work, we advocate for viewing the prediction problem through the lens of compression and evaluate the compression capabilities of large (foundation) models. We show that large language models are p"},"claims":{"count":4,"items":[{"kind":"strongest_claim","text":"Chinchilla 70B, while trained primarily on text, compresses ImageNet patches to 43.4% and LibriSpeech samples to 16.4% of their raw size, beating domain-specific compressors like PNG (58.5%) or FLAC (30.3%), respectively.","source":"verdict.strongest_claim","status":"machine_extracted","claim_id":"C1","attestation":"unclaimed"},{"kind":"weakest_assumption","text":"That the predictive distribution produced by the language model can be directly converted into a lossless compression scheme via arithmetic coding without significant overhead or implementation-specific losses that would invalidate the reported ratios.","source":"verdict.weakest_assumption","status":"machine_extracted","claim_id":"C2","attestation":"unclaimed"},{"kind":"one_line_summary","text":"Large language models serve as strong general-purpose lossless compressors for text, images, and audio, outperforming domain-specific methods and revealing insights into scaling, tokenization, and in-context learning.","source":"verdict.one_line_summary","status":"machine_extracted","claim_id":"C3","attestation":"unclaimed"},{"kind":"headline","text":"Large language models trained on text compress images and audio better than specialized tools.","source":"verdict.pith_extraction.headline","status":"machine_extracted","claim_id":"C4","attestation":"unclaimed"}],"snapshot_sha256":"3127474625bf7df2d66d51df773744c43db31e0d5e7b06b7bcf88b9fc8c92bdd"},"source":{"id":"2309.10668","kind":"arxiv","version":2},"verdict":{"id":"4f9f2a55-2aa3-4a8d-bd72-9de4fb57a489","model_set":{"reader":"grok-4.3"},"created_at":"2026-05-17T22:31:39.076349Z","strongest_claim":"Chinchilla 70B, while trained primarily on text, compresses ImageNet patches to 43.4% and LibriSpeech samples to 16.4% of their raw size, beating domain-specific compressors like PNG (58.5%) or FLAC (30.3%), respectively.","one_line_summary":"Large language models serve as strong general-purpose lossless compressors for text, images, and audio, outperforming domain-specific methods and revealing insights into scaling, tokenization, and in-context learning.","pipeline_version":"pith-pipeline@v0.9.0","weakest_assumption":"That the predictive distribution produced by the language model can be directly converted into a lossless compression scheme via arithmetic coding without significant overhead or implementation-specific losses that would invalidate the reported ratios.","pith_extraction_headline":"Large language models trained on text compress images and audio better than specialized tools."},"references":{"count":20,"sample":[{"doi":"","year":null,"title":"On the Opportunities and Risks of Foundation Models","work_id":"a18039e9-928d-47c9-a836-32656a71bf71","ref_index":1,"cited_arxiv_id":"2108.07258","is_internal_anchor":true},{"doi":"","year":null,"title":"Sparks of Artificial General Intelligence: Early experiments with GPT-4","work_id":"a23cfe92-7f7c-424b-98d4-b386a83002fb","ref_index":2,"cited_arxiv_id":"2303.12712","is_internal_anchor":true},{"doi":"","year":null,"title":"Scaling transformer to 1m tokens and beyond with rmt","work_id":"f5b46c3e-9f26-481f-a773-689cbd972ec4","ref_index":3,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":null,"title":"arXiv preprint arXiv:1710.09282 , year=","work_id":"95630094-c860-4769-b4d5-042827b8f21e","ref_index":4,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":null,"title":"Syntactically Informed Text Compression with Recurrent Neural Networks","work_id":"d37d28de-8bf2-46c9-8cc2-5434dd452dba","ref_index":5,"cited_arxiv_id":"1608.02893","is_internal_anchor":true}],"resolved_work":20,"snapshot_sha256":"85e30d7099f817ef42e31319b401db64fea872e28abdcd5950d96918fd90f139","internal_anchors":8},"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":"2309.10668","created_at":"2026-05-17T23:38:12.795277+00:00"},{"alias_kind":"arxiv_version","alias_value":"2309.10668v2","created_at":"2026-05-17T23:38:12.795277+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2309.10668","created_at":"2026-05-17T23:38:12.795277+00:00"},{"alias_kind":"pith_short_12","alias_value":"XTLBBB6Y4KTU","created_at":"2026-05-18T12:33:37.589309+00:00"},{"alias_kind":"pith_short_16","alias_value":"XTLBBB6Y4KTUBXU5","created_at":"2026-05-18T12:33:37.589309+00:00"},{"alias_kind":"pith_short_8","alias_value":"XTLBBB6Y","created_at":"2026-05-18T12:33:37.589309+00:00"}],"events":[],"event_summary":{},"paper_claims":[],"inbound_citations":{"count":18,"internal_anchor_count":18,"sample":[{"citing_arxiv_id":"2505.17469","citing_title":"Efficient compression of neural networks and datasets","ref_index":11,"is_internal_anchor":true},{"citing_arxiv_id":"2601.20255","citing_title":"HE-SNR: Uncovering Latent Logic via Entropy for Guiding Mid-Training on SWE-bench","ref_index":5,"is_internal_anchor":true},{"citing_arxiv_id":"2406.20094","citing_title":"Scaling Synthetic Data Creation with 1,000,000,000 Personas","ref_index":10,"is_internal_anchor":true},{"citing_arxiv_id":"2605.05209","citing_title":"Are Flat Minima an Illusion?","ref_index":160,"is_internal_anchor":true},{"citing_arxiv_id":"2604.06228","citing_title":"Probabilistic Language Tries: A Unified Framework for Compression, Decision Policies, and Execution Reuse","ref_index":4,"is_internal_anchor":true},{"citing_arxiv_id":"2605.13485","citing_title":"Effective Context in Transformers: An Analysis of Fragmentation and Tokenization","ref_index":10,"is_internal_anchor":true},{"citing_arxiv_id":"2311.05232","citing_title":"A Survey on Hallucination in Large Language Models: Principles, Taxonomy, Challenges, and Open Questions","ref_index":74,"is_internal_anchor":true},{"citing_arxiv_id":"2604.18486","citing_title":"Xiaomi OneVL: One-Step Latent Reasoning and Planning with Vision-Language Explanation","ref_index":20,"is_internal_anchor":true},{"citing_arxiv_id":"2605.09724","citing_title":"Model Capacity Determines Grokking through Competing Memorisation and Generalisation Speeds","ref_index":8,"is_internal_anchor":true},{"citing_arxiv_id":"2605.05115","citing_title":"Manifold Steering Reveals the Shared Geometry of Neural Network Representation and Behavior","ref_index":248,"is_internal_anchor":true},{"citing_arxiv_id":"2605.01973","citing_title":"Learn-to-learn on Arbitrary Textual Conditioning: A Hypernetwork-Driven Meta-Gated LLM","ref_index":148,"is_internal_anchor":true},{"citing_arxiv_id":"2404.18930","citing_title":"Hallucination of Multimodal Large Language Models: A Survey","ref_index":40,"is_internal_anchor":true},{"citing_arxiv_id":"2604.15356","citing_title":"Sequential KV Cache Compression via Probabilistic Language Tries: Beyond the Per-Vector Shannon Limit","ref_index":3,"is_internal_anchor":true},{"citing_arxiv_id":"2604.06754","citing_title":"The Rhetoric of Machine Learning","ref_index":1,"is_internal_anchor":true},{"citing_arxiv_id":"2604.07239","citing_title":"Efficient Learned Data Compression via Dual-Stream Feature Decoupling","ref_index":1,"is_internal_anchor":true},{"citing_arxiv_id":"2604.05072","citing_title":"Hierarchical SVG Tokenization: Learning Compact Visual Programs for Scalable Vector Graphics Modeling","ref_index":6,"is_internal_anchor":true},{"citing_arxiv_id":"2604.15472","citing_title":"Lossless Compression via Chained Lightweight Neural Predictors with Information Inheritance","ref_index":6,"is_internal_anchor":true},{"citing_arxiv_id":"2604.18486","citing_title":"Xiaomi OneVL: One-Step Latent Reasoning and Planning with Vision-Language Explanation","ref_index":20,"is_internal_anchor":true}]},"formal_canon":{"evidence_count":0,"sample":[],"anchors":[]},"links":{"html":"https://pith.science/pith/XTLBBB6Y4KTUBXU5EM27RGCOF7","json":"https://pith.science/pith/XTLBBB6Y4KTUBXU5EM27RGCOF7.json","graph_json":"https://pith.science/api/pith-number/XTLBBB6Y4KTUBXU5EM27RGCOF7/graph.json","events_json":"https://pith.science/api/pith-number/XTLBBB6Y4KTUBXU5EM27RGCOF7/events.json","paper":"https://pith.science/paper/XTLBBB6Y"},"agent_actions":{"view_html":"https://pith.science/pith/XTLBBB6Y4KTUBXU5EM27RGCOF7","download_json":"https://pith.science/pith/XTLBBB6Y4KTUBXU5EM27RGCOF7.json","view_paper":"https://pith.science/paper/XTLBBB6Y","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2309.10668&json=true","fetch_graph":"https://pith.science/api/pith-number/XTLBBB6Y4KTUBXU5EM27RGCOF7/graph.json","fetch_events":"https://pith.science/api/pith-number/XTLBBB6Y4KTUBXU5EM27RGCOF7/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/XTLBBB6Y4KTUBXU5EM27RGCOF7/action/timestamp_anchor","attest_storage":"https://pith.science/pith/XTLBBB6Y4KTUBXU5EM27RGCOF7/action/storage_attestation","attest_author":"https://pith.science/pith/XTLBBB6Y4KTUBXU5EM27RGCOF7/action/author_attestation","sign_citation":"https://pith.science/pith/XTLBBB6Y4KTUBXU5EM27RGCOF7/action/citation_signature","submit_replication":"https://pith.science/pith/XTLBBB6Y4KTUBXU5EM27RGCOF7/action/replication_record"}},"created_at":"2026-05-17T23:38:12.795277+00:00","updated_at":"2026-05-17T23:38:12.795277+00:00"}