{"paper":{"title":"The Efficiency Gap in Byte Modeling","license":"http://creativecommons.org/licenses/by/4.0/","headline":"Byte modeling incurs a larger scaling penalty under masked diffusion than under autoregressive training because diffusion destroys local byte contiguity.","cross_cats":[],"primary_cat":"cs.LG","authors_text":"Alexander M. Rush, Celine Lee, Chen Liang, Derek Cheng, Ed Chi, Fernando Pereira, Jeremiah Liu, Jiaxin Shi, Jing Nathan Yan, Pengcheng Yin, Ruoxi Wang, Yin Zhang","submitted_at":"2026-05-13T03:03:30Z","abstract_excerpt":"Modern language models have historically relied on two dominant design choices: subword tokenization and autoregressive (AR) ordering. These design decisions bake in priors that dictate a model's learning. Recently, two alternative paradigms have challenged this: byte-level modeling, which bypasses static statistically-derived token vocabularies, and masked diffusion modeling (MDM), which conducts parallel, non-sequential generation. Their intersection represents a fully end-to-end modality-agnostic generative prototype; however, removing these structural priors incurs a significant computatio"},"claims":{"count":4,"items":[{"kind":"strongest_claim","text":"the performance penalty of byte modeling is not uniform; across scale, the scaling overhead of byte modeling is worse for MDM than for AR. We hypothesize that this disparity stems from context fragility: while AR's stable causal history allows models to naturally rediscover subword patterns, the MDM objective destroys the local contiguity required to efficiently resolve semantics from raw bytes.","source":"verdict.strongest_claim","status":"machine_extracted","claim_id":"C1","attestation":"unclaimed"},{"kind":"weakest_assumption","text":"that the observed scaling disparity is caused by context fragility in MDM rather than differences in how compute is allocated or other unmeasured factors in the experimental setup.","source":"verdict.weakest_assumption","status":"machine_extracted","claim_id":"C2","attestation":"unclaimed"},{"kind":"one_line_summary","text":"Byte modeling incurs greater scaling overhead for masked diffusion than autoregressive models because the diffusion objective destroys local byte contiguity needed to resolve semantics.","source":"verdict.one_line_summary","status":"machine_extracted","claim_id":"C3","attestation":"unclaimed"},{"kind":"headline","text":"Byte modeling incurs a larger scaling penalty under masked diffusion than under autoregressive training because diffusion destroys local byte contiguity.","source":"verdict.pith_extraction.headline","status":"machine_extracted","claim_id":"C4","attestation":"unclaimed"}],"snapshot_sha256":"91b77bcb96316b90d3fb483bee7bd37220fbc0f4c4b31f257ec78b0cf6f135af"},"source":{"id":"2605.12928","kind":"arxiv","version":1},"verdict":{"id":"270adfc7-ae15-4a01-a746-3aaba57b3726","model_set":{"reader":"grok-4.3"},"created_at":"2026-05-14T19:40:05.302095Z","strongest_claim":"the performance penalty of byte modeling is not uniform; across scale, the scaling overhead of byte modeling is worse for MDM than for AR. We hypothesize that this disparity stems from context fragility: while AR's stable causal history allows models to naturally rediscover subword patterns, the MDM objective destroys the local contiguity required to efficiently resolve semantics from raw bytes.","one_line_summary":"Byte modeling incurs greater scaling overhead for masked diffusion than autoregressive models because the diffusion objective destroys local byte contiguity needed to resolve semantics.","pipeline_version":"pith-pipeline@v0.9.0","weakest_assumption":"that the observed scaling disparity is caused by context fragility in MDM rather than differences in how compute is allocated or other unmeasured factors in the experimental setup.","pith_extraction_headline":"Byte modeling incurs a larger scaling penalty under masked diffusion than under autoregressive training because diffusion destroys local byte contiguity."},"references":{"count":51,"sample":[{"doi":"","year":2025,"title":"Adapters for Altering","work_id":"e8429366-c015-4c81-b8b7-42fc056c1caf","ref_index":1,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":null,"title":"Arnaud Pannatier and Evann Courdier and François Fleuret , year=. 2404.09562 , archivePrefix=","work_id":"98f61b5e-922e-40ff-b51f-f75fe5c9b44a","ref_index":2,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":null,"title":"Shkarin, D. A. , title =. Problems of Information Transmission , year =","work_id":"28e15183-5ccc-4733-9f87-281db91d5733","ref_index":3,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2024,"title":"2024 , howpublished =","work_id":"f88ce759-208c-4f98-8571-9cdded5ee014","ref_index":4,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2021,"title":"Kucherawy , title =","work_id":"78597206-9893-4182-abb6-fb475f931ef5","ref_index":5,"cited_arxiv_id":"","is_internal_anchor":false}],"resolved_work":51,"snapshot_sha256":"1155c111f33740be3ecb76eeed05967e4c78ad545e13729368ada515c0e372da","internal_anchors":3},"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"}