{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2025:DTTEN7LW6HIIOJ22TOQV4HOQWO","short_pith_number":"pith:DTTEN7LW","schema_version":"1.0","canonical_sha256":"1ce646fd76f1d087275a9ba15e1dd0b3bfe69d47a172bf55dd0e4acef05bba12","source":{"kind":"arxiv","id":"2502.15955","version":1},"attestation_state":"computed","paper":{"title":"Compression Barriers for Autoregressive Transformers","license":"http://creativecommons.org/licenses/by/4.0/","headline":"","cross_cats":["cs.AI","cs.CC","cs.LG"],"primary_cat":"cs.DS","authors_text":"Krzysztof Onak, Themistoklis Haris","submitted_at":"2025-02-21T21:37:52Z","abstract_excerpt":"A key limitation of autoregressive Transformers is the large memory needed at inference-time to cache all previous key-value (KV) embeddings. Prior works address this by compressing the KV cache, but often assume specific structural properties of the embeddings. This raises the following natural question: Can truly sublinear space utilization be achieved without such assumptions? In this work, we answer this question in the negative. Any algorithm for attention-based token generation must use $\\Theta(nd)$ space, where $n$ is the number of tokens generated so far and $d = \\Omega(\\log n)$ is the"},"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":"2502.15955","kind":"arxiv","version":1},"metadata":{"license":"http://creativecommons.org/licenses/by/4.0/","primary_cat":"cs.DS","submitted_at":"2025-02-21T21:37:52Z","cross_cats_sorted":["cs.AI","cs.CC","cs.LG"],"title_canon_sha256":"91d9cce554097163955cde3de3b4237b77143571d41f26614a72318a460307f6","abstract_canon_sha256":"217af31f7ea621f6a0bcf2d5f7a74b663a9b1656b1649ef2aad1678fbb1af686"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-07-05T10:18:19.010645Z","signature_b64":"46yT0393OpuPFQqLpTIlQ/fCyXPkoa8H3ShuD10Sar4Y2ZWVRQAPYwZ/xEK7X5Ofp2pYX9NXbVVRXAQdX9JYCg==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"1ce646fd76f1d087275a9ba15e1dd0b3bfe69d47a172bf55dd0e4acef05bba12","last_reissued_at":"2026-07-05T10:18:19.010171Z","signature_status":"signed_v1","first_computed_at":"2026-07-05T10:18:19.010171Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"Compression Barriers for Autoregressive Transformers","license":"http://creativecommons.org/licenses/by/4.0/","headline":"","cross_cats":["cs.AI","cs.CC","cs.LG"],"primary_cat":"cs.DS","authors_text":"Krzysztof Onak, Themistoklis Haris","submitted_at":"2025-02-21T21:37:52Z","abstract_excerpt":"A key limitation of autoregressive Transformers is the large memory needed at inference-time to cache all previous key-value (KV) embeddings. Prior works address this by compressing the KV cache, but often assume specific structural properties of the embeddings. This raises the following natural question: Can truly sublinear space utilization be achieved without such assumptions? In this work, we answer this question in the negative. Any algorithm for attention-based token generation must use $\\Theta(nd)$ space, where $n$ is the number of tokens generated so far and $d = \\Omega(\\log n)$ is the"},"claims":{"count":0,"items":[],"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"source":{"id":"2502.15955","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/2502.15955/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":"2502.15955","created_at":"2026-07-05T10:18:19.010240+00:00"},{"alias_kind":"arxiv_version","alias_value":"2502.15955v1","created_at":"2026-07-05T10:18:19.010240+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2502.15955","created_at":"2026-07-05T10:18:19.010240+00:00"},{"alias_kind":"pith_short_12","alias_value":"DTTEN7LW6HII","created_at":"2026-07-05T10:18:19.010240+00:00"},{"alias_kind":"pith_short_16","alias_value":"DTTEN7LW6HIIOJ22","created_at":"2026-07-05T10:18:19.010240+00:00"},{"alias_kind":"pith_short_8","alias_value":"DTTEN7LW","created_at":"2026-07-05T10:18:19.010240+00:00"}],"events":[],"event_summary":{},"paper_claims":[],"inbound_citations":{"count":3,"internal_anchor_count":1,"sample":[{"citing_arxiv_id":"2607.08032","citing_title":"What to Keep, What to Forget: A Rate--Distortion View of Memory Compaction in LLMs and Agents","ref_index":48,"is_internal_anchor":true},{"citing_arxiv_id":"2607.01520","citing_title":"The risk of KV cache compression","ref_index":16,"is_internal_anchor":false},{"citing_arxiv_id":"2604.17935","citing_title":"How Much Cache Does Reasoning Need? Depth-Cache Tradeoffs in KV-Compressed Transformers","ref_index":6,"is_internal_anchor":false}]},"formal_canon":{"evidence_count":0,"sample":[],"anchors":[]},"links":{"html":"https://pith.science/pith/DTTEN7LW6HIIOJ22TOQV4HOQWO","json":"https://pith.science/pith/DTTEN7LW6HIIOJ22TOQV4HOQWO.json","graph_json":"https://pith.science/api/pith-number/DTTEN7LW6HIIOJ22TOQV4HOQWO/graph.json","events_json":"https://pith.science/api/pith-number/DTTEN7LW6HIIOJ22TOQV4HOQWO/events.json","paper":"https://pith.science/paper/DTTEN7LW"},"agent_actions":{"view_html":"https://pith.science/pith/DTTEN7LW6HIIOJ22TOQV4HOQWO","download_json":"https://pith.science/pith/DTTEN7LW6HIIOJ22TOQV4HOQWO.json","view_paper":"https://pith.science/paper/DTTEN7LW","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2502.15955&json=true","fetch_graph":"https://pith.science/api/pith-number/DTTEN7LW6HIIOJ22TOQV4HOQWO/graph.json","fetch_events":"https://pith.science/api/pith-number/DTTEN7LW6HIIOJ22TOQV4HOQWO/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/DTTEN7LW6HIIOJ22TOQV4HOQWO/action/timestamp_anchor","attest_storage":"https://pith.science/pith/DTTEN7LW6HIIOJ22TOQV4HOQWO/action/storage_attestation","attest_author":"https://pith.science/pith/DTTEN7LW6HIIOJ22TOQV4HOQWO/action/author_attestation","sign_citation":"https://pith.science/pith/DTTEN7LW6HIIOJ22TOQV4HOQWO/action/citation_signature","submit_replication":"https://pith.science/pith/DTTEN7LW6HIIOJ22TOQV4HOQWO/action/replication_record"}},"created_at":"2026-07-05T10:18:19.010240+00:00","updated_at":"2026-07-05T10:18:19.010240+00:00"}