{"paper":{"title":"Large Language Model as Token Compressor and Decompressor","license":"http://creativecommons.org/licenses/by/4.0/","headline":"An off-the-shelf LLM can be fine-tuned with LoRA to compress long texts into adaptive sequences of Z-tokens while preserving reconstruction and task performance.","cross_cats":[],"primary_cat":"cs.CL","authors_text":"Jielei Zhang, Junkai Lin, Tianhao Zhao, Wei Yang, Wenbing Li, Yiran Wang, Zikai Song","submitted_at":"2026-03-26T11:30:44Z","abstract_excerpt":"In this paper, we study whether an off-the-shelf LLM can be adapted into a discrete, variable-length token compressor and decompressor for long-context processing. To this end, we design a self-expressive autoencoding framework that fine-tunes a pretrained LLM with lightweight LoRA adapters to map long texts into compact sequences of learned latent codes, termed Z-tokens, and to decode them back into natural language or task outputs. The resulting representation is content-adaptive: less predictable or information-dense segments can receive more Z-tokens, while redundant regions can be represe"},"claims":{"count":4,"items":[{"kind":"strongest_claim","text":"an off-the-shelf LLM can be adapted into a discrete, variable-length token compressor and decompressor for long-context processing","source":"verdict.strongest_claim","status":"machine_extracted","claim_id":"C1","attestation":"unclaimed"},{"kind":"weakest_assumption","text":"that fine-tuning with LoRA on the self-expressive autoencoding objective will produce Z-tokens that preserve enough information for both faithful reconstruction and downstream task performance without requiring extensive post-hoc adjustments","source":"verdict.weakest_assumption","status":"machine_extracted","claim_id":"C2","attestation":"unclaimed"},{"kind":"one_line_summary","text":"A pretrained LLM is adapted via LoRA fine-tuning into a content-adaptive compressor that maps long texts to compact variable-length Z-token sequences while preserving reconstruction quality and downstream performance.","source":"verdict.one_line_summary","status":"machine_extracted","claim_id":"C3","attestation":"unclaimed"},{"kind":"headline","text":"An off-the-shelf LLM can be fine-tuned with LoRA to compress long texts into adaptive sequences of Z-tokens while preserving reconstruction and task performance.","source":"verdict.pith_extraction.headline","status":"machine_extracted","claim_id":"C4","attestation":"unclaimed"}],"snapshot_sha256":"ab74a9d5a7564ab3a0b542b623027b4812b61d5a48311c23e4afd59113ab3f18"},"source":{"id":"2603.25340","kind":"arxiv","version":2},"verdict":{"id":"aeeb3b15-b7ea-4d5e-b3d9-aa7f68a69caf","model_set":{"reader":"grok-4.3"},"created_at":"2026-05-15T00:51:31.603008Z","strongest_claim":"an off-the-shelf LLM can be adapted into a discrete, variable-length token compressor and decompressor for long-context processing","one_line_summary":"A pretrained LLM is adapted via LoRA fine-tuning into a content-adaptive compressor that maps long texts to compact variable-length Z-token sequences while preserving reconstruction quality and downstream performance.","pipeline_version":"pith-pipeline@v0.9.0","weakest_assumption":"that fine-tuning with LoRA on the self-expressive autoencoding objective will produce Z-tokens that preserve enough information for both faithful reconstruction and downstream task performance without requiring extensive post-hoc adjustments","pith_extraction_headline":"An off-the-shelf LLM can be fine-tuned with LoRA to compress long texts into adaptive sequences of Z-tokens while preserving reconstruction and task performance."},"references":{"count":47,"sample":[{"doi":"","year":2020,"title":"Peters, and Arman Cohan","work_id":"477d1550-5247-4659-85a2-1cc846c63115","ref_index":1,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2023,"title":"Token merging: Your vit but faster, 2023","work_id":"790954c9-35d8-4397-92ae-24d94c563cc4","ref_index":2,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2022,"title":"Hudson, Ehsan Adeli, et al","work_id":"10e2738f-b505-4455-824c-58c06c17d125","ref_index":3,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2020,"title":"Tom B. Brown, Benjamin Mann, Nick Ryder, Melanie Sub- biah, Jared Kaplan, Prafulla Dhariwal, Arvind Neelakan- tan, Pranav Shyam, Girish Sastry, Amanda Askell, Sand- hini Agarwal, Ariel Herbert-V oss, ","work_id":"817f656c-a9d7-4ab2-a333-8160212f136f","ref_index":4,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2023,"title":"Adapting language models to compress con- texts, 2023","work_id":"b022c61c-4cd4-40d2-b589-dbb244102410","ref_index":5,"cited_arxiv_id":"","is_internal_anchor":false}],"resolved_work":47,"snapshot_sha256":"2692c8234c92c8b3dd6d8180afe3b3c8ef62bd5294d4beb60c953b75c416b154","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"}