{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2025:X77LESH6A34BJBQS4HXB3YXMXU","short_pith_number":"pith:X77LESH6","schema_version":"1.0","canonical_sha256":"bffeb248fe06f8148612e1ee1de2ecbd2f934a0472556aa67611fe3d75503f94","source":{"kind":"arxiv","id":"2506.14761","version":1},"attestation_state":"computed","paper":{"title":"From Bytes to Ideas: Language Modeling with Autoregressive U-Nets","license":"http://creativecommons.org/licenses/by/4.0/","headline":"","cross_cats":["cs.AI"],"primary_cat":"cs.CL","authors_text":"Alessandro Leite, Badr Youbi Idrissi, David Lopez-Paz, Marc Schoenauer, Mathurin Videau, Olivier Teytaud","submitted_at":"2025-06-17T17:55:11Z","abstract_excerpt":"Tokenization imposes a fixed granularity on the input text, freezing how a language model operates on data and how far in the future it predicts. Byte Pair Encoding (BPE) and similar schemes split text once, build a static vocabulary, and leave the model stuck with that choice. We relax this rigidity by introducing an autoregressive U-Net that learns to embed its own tokens as it trains. The network reads raw bytes, pools them into words, then pairs of words, then up to 4 words, giving it a multi-scale view of the sequence. At deeper stages, the model must predict further into the future -- an"},"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":"2506.14761","kind":"arxiv","version":1},"metadata":{"license":"http://creativecommons.org/licenses/by/4.0/","primary_cat":"cs.CL","submitted_at":"2025-06-17T17:55:11Z","cross_cats_sorted":["cs.AI"],"title_canon_sha256":"1fca76a35686ecd4559b3f9e9655759add15c6786966825a19cce762a2edbf66","abstract_canon_sha256":"82bc1477f53cc59538e4b6f8920ef18837eccd53f053aa0da479bd381bf72273"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-07-05T11:23:12.516154Z","signature_b64":"Vusu/lI6O5G0G5Vmi4iLX+9A6dS1bjC9bXJNuxkDaw2yCZACfsW4f0A3R1RQxp1a3cRT4UEAb6QOn+BYT1QdCg==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"bffeb248fe06f8148612e1ee1de2ecbd2f934a0472556aa67611fe3d75503f94","last_reissued_at":"2026-07-05T11:23:12.515524Z","signature_status":"signed_v1","first_computed_at":"2026-07-05T11:23:12.515524Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"From Bytes to Ideas: Language Modeling with Autoregressive U-Nets","license":"http://creativecommons.org/licenses/by/4.0/","headline":"","cross_cats":["cs.AI"],"primary_cat":"cs.CL","authors_text":"Alessandro Leite, Badr Youbi Idrissi, David Lopez-Paz, Marc Schoenauer, Mathurin Videau, Olivier Teytaud","submitted_at":"2025-06-17T17:55:11Z","abstract_excerpt":"Tokenization imposes a fixed granularity on the input text, freezing how a language model operates on data and how far in the future it predicts. Byte Pair Encoding (BPE) and similar schemes split text once, build a static vocabulary, and leave the model stuck with that choice. We relax this rigidity by introducing an autoregressive U-Net that learns to embed its own tokens as it trains. The network reads raw bytes, pools them into words, then pairs of words, then up to 4 words, giving it a multi-scale view of the sequence. At deeper stages, the model must predict further into the future -- an"},"claims":{"count":0,"items":[],"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"source":{"id":"2506.14761","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/2506.14761/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":"2506.14761","created_at":"2026-07-05T11:23:12.515630+00:00"},{"alias_kind":"arxiv_version","alias_value":"2506.14761v1","created_at":"2026-07-05T11:23:12.515630+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2506.14761","created_at":"2026-07-05T11:23:12.515630+00:00"},{"alias_kind":"pith_short_12","alias_value":"X77LESH6A34B","created_at":"2026-07-05T11:23:12.515630+00:00"},{"alias_kind":"pith_short_16","alias_value":"X77LESH6A34BJBQS","created_at":"2026-07-05T11:23:12.515630+00:00"},{"alias_kind":"pith_short_8","alias_value":"X77LESH6","created_at":"2026-07-05T11:23:12.515630+00:00"}],"events":[],"event_summary":{},"paper_claims":[],"inbound_citations":{"count":5,"internal_anchor_count":0,"sample":[{"citing_arxiv_id":"2606.28057","citing_title":"MultiHashFormer: Hash-based Generative Language Models","ref_index":50,"is_internal_anchor":false},{"citing_arxiv_id":"2602.04289","citing_title":"Proxy Compression for Language Modeling","ref_index":13,"is_internal_anchor":false},{"citing_arxiv_id":"2605.09630","citing_title":"Scratchpad Patching: Decoupling Compute from Patch Size in Byte-Level Language Models","ref_index":89,"is_internal_anchor":false},{"citing_arxiv_id":"2605.01188","citing_title":"Compute Optimal Tokenization","ref_index":34,"is_internal_anchor":false},{"citing_arxiv_id":"2604.07466","citing_title":"Cross-Tokenizer LLM Distillation through a Byte-Level Interface","ref_index":1,"is_internal_anchor":false}]},"formal_canon":{"evidence_count":0,"sample":[],"anchors":[]},"links":{"html":"https://pith.science/pith/X77LESH6A34BJBQS4HXB3YXMXU","json":"https://pith.science/pith/X77LESH6A34BJBQS4HXB3YXMXU.json","graph_json":"https://pith.science/api/pith-number/X77LESH6A34BJBQS4HXB3YXMXU/graph.json","events_json":"https://pith.science/api/pith-number/X77LESH6A34BJBQS4HXB3YXMXU/events.json","paper":"https://pith.science/paper/X77LESH6"},"agent_actions":{"view_html":"https://pith.science/pith/X77LESH6A34BJBQS4HXB3YXMXU","download_json":"https://pith.science/pith/X77LESH6A34BJBQS4HXB3YXMXU.json","view_paper":"https://pith.science/paper/X77LESH6","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2506.14761&json=true","fetch_graph":"https://pith.science/api/pith-number/X77LESH6A34BJBQS4HXB3YXMXU/graph.json","fetch_events":"https://pith.science/api/pith-number/X77LESH6A34BJBQS4HXB3YXMXU/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/X77LESH6A34BJBQS4HXB3YXMXU/action/timestamp_anchor","attest_storage":"https://pith.science/pith/X77LESH6A34BJBQS4HXB3YXMXU/action/storage_attestation","attest_author":"https://pith.science/pith/X77LESH6A34BJBQS4HXB3YXMXU/action/author_attestation","sign_citation":"https://pith.science/pith/X77LESH6A34BJBQS4HXB3YXMXU/action/citation_signature","submit_replication":"https://pith.science/pith/X77LESH6A34BJBQS4HXB3YXMXU/action/replication_record"}},"created_at":"2026-07-05T11:23:12.515630+00:00","updated_at":"2026-07-05T11:23:12.515630+00:00"}