{"paper":{"title":"Proxy Compression for Language Modeling","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","headline":"Proxy compression trains language models jointly on raw bytes and compressed sequences so they can use efficient inputs during training yet run purely on raw bytes at inference.","cross_cats":["cs.LG"],"primary_cat":"cs.CL","authors_text":"Lingpeng Kong, Lin Zheng, Qian Liu, Xiachong Feng, Xinyu Li","submitted_at":"2026-02-04T07:36:46Z","abstract_excerpt":"Modern language models are trained almost exclusively on token sequences produced by a fixed tokenizer, an external lossless compressor often over UTF-8 byte sequences, thereby coupling the model to that compressor. This work introduces proxy compression, an alternative training scheme that preserves the efficiency benefits of compressed inputs while providing an end-to-end, raw-byte interface at inference time. During training, a single language model is jointly trained on raw byte sequences and compressed views generated by external compressors; through the process, the model learns to inter"},"claims":{"count":4,"items":[{"kind":"strongest_claim","text":"proxy compression substantially improves training efficiency and significantly outperforms pure byte-level baselines given fixed compute budgets. As model scale increases, these gains become more pronounced, and proxy-trained models eventually match or surpass tokenizer approaches, all while operating solely on raw bytes.","source":"verdict.strongest_claim","status":"machine_extracted","claim_id":"C1","attestation":"unclaimed"},{"kind":"weakest_assumption","text":"that the internal alignment learned during joint training on compressed and raw views transfers effectively to pure raw-byte inference without performance degradation or the need for continued compressed inputs.","source":"verdict.weakest_assumption","status":"machine_extracted","claim_id":"C2","attestation":"unclaimed"},{"kind":"one_line_summary","text":"Proxy compression trains language models on both raw bytes and compressed sequences to enable efficient training with raw-byte inference at test time.","source":"verdict.one_line_summary","status":"machine_extracted","claim_id":"C3","attestation":"unclaimed"},{"kind":"headline","text":"Proxy compression trains language models jointly on raw bytes and compressed sequences so they can use efficient inputs during training yet run purely on raw bytes at inference.","source":"verdict.pith_extraction.headline","status":"machine_extracted","claim_id":"C4","attestation":"unclaimed"}],"snapshot_sha256":"bd42ab1bbd133da7cccab68460c3a479d334099daa7df374cf33c9228707aed3"},"source":{"id":"2602.04289","kind":"arxiv","version":2},"verdict":{"id":"8f795edd-4a58-4c2c-845f-4e2ba09840a1","model_set":{"reader":"grok-4.3"},"created_at":"2026-05-16T08:00:32.518982Z","strongest_claim":"proxy compression substantially improves training efficiency and significantly outperforms pure byte-level baselines given fixed compute budgets. As model scale increases, these gains become more pronounced, and proxy-trained models eventually match or surpass tokenizer approaches, all while operating solely on raw bytes.","one_line_summary":"Proxy compression trains language models on both raw bytes and compressed sequences to enable efficient training with raw-byte inference at test time.","pipeline_version":"pith-pipeline@v0.9.0","weakest_assumption":"that the internal alignment learned during joint training on compressed and raw views transfers effectively to pure raw-byte inference without performance degradation or the need for continued compressed inputs.","pith_extraction_headline":"Proxy compression trains language models jointly on raw bytes and compressed sequences so they can use efficient inputs during training yet run purely on raw bytes at inference."},"references":{"count":29,"sample":[{"doi":"","year":2023,"title":"Magnet: Improving the multilingual fairness of language models with adaptive gradient-based tokenization","work_id":"c19a5144-aef7-4515-a0d5-05fde00aa9e8","ref_index":1,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2021,"title":"URL https://openreview.net/forum? id=PEpbUobfJv. Cao, K. and Rimell, L. You should evaluate your language model on marginal likelihood over tokeni- sations. InProceedings of the 2021 Conference on Emp","work_id":"a9c0a510-0c8c-4107-a1a5-bb766e8da0c9","ref_index":2,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2021,"title":"Evaluating Large Language Models Trained on Code","work_id":"042493e9-b26f-4b4e-bbde-382072ca9b08","ref_index":3,"cited_arxiv_id":"2107.03374","is_internal_anchor":true},{"doi":"10.18653/v1/n19-1423","year":2019,"title":"The Llama 3 Herd of Models","work_id":"1549a635-88af-4ac1-acfe-51ae7bb53345","ref_index":4,"cited_arxiv_id":"2407.21783","is_internal_anchor":true},{"doi":"","year":2023,"title":"emnlp-industry.58/","work_id":"93368c2d-5c7f-41ec-a047-47b30f46d1ae","ref_index":5,"cited_arxiv_id":"","is_internal_anchor":false}],"resolved_work":29,"snapshot_sha256":"41949df9a87d04c09e9a5d873d7ffe1991c9e5f2faee63e83e5ae8c1a4c0379b","internal_anchors":4},"formal_canon":{"evidence_count":2,"snapshot_sha256":"89d9492ac551897fc9375862f20ae303877ad7b91e984ff059817624ff94b4de"},"author_claims":{"count":0,"strong_count":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"builder_version":"pith-number-builder-2026-05-17-v1"}