{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2026:VJAH75QSESYAYAT5N7M3YXPYYD","short_pith_number":"pith:VJAH75QS","schema_version":"1.0","canonical_sha256":"aa407ff61224b00c027d6fd9bc5df8c0d85c5e2b0ccd81c7872d5d01837ef41b","source":{"kind":"arxiv","id":"2602.10238","version":2},"attestation_state":"computed","paper":{"title":"Learning to Evict from Key-Value Cache","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","headline":"","cross_cats":["cs.LG"],"primary_cat":"cs.CL","authors_text":"Laura Manduchi, Luca Moschella, Ozan Sener","submitted_at":"2026-02-10T19:34:15Z","abstract_excerpt":"The growing size of Large Language Models (LLMs) makes efficient inference challenging, primarily due to the memory demands of the autoregressive Key-Value (KV) cache. Existing eviction or compression methods reduce cost but rely on heuristics, such as recency or past attention scores, which serve only as indirect proxies for a token's future utility and introduce computational overhead. We reframe KV cache eviction as a reinforcement learning (RL) problem: learning to rank tokens by their predicted usefulness for future decoding. To this end, we introduce KV Policy (KVP), a framework of light"},"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":"2602.10238","kind":"arxiv","version":2},"metadata":{"license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","primary_cat":"cs.CL","submitted_at":"2026-02-10T19:34:15Z","cross_cats_sorted":["cs.LG"],"title_canon_sha256":"1b89f3b3ed46f747d0c9267565474b280dd2e42026cd9c650c98781702f83880","abstract_canon_sha256":"abe7916928050c861c07d9702d445db923bbaece126744021d0da2820c830c48"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-06-29T01:14:28.827449Z","signature_b64":"X5guVQoziU5Ja4bzN3j327JqK5//D0LTjMNB9jLKEAto6IYcCNfBAsb2cSuE5SMWaLTclS+8HuhiIzbi8nmfAA==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"aa407ff61224b00c027d6fd9bc5df8c0d85c5e2b0ccd81c7872d5d01837ef41b","last_reissued_at":"2026-06-29T01:14:28.826900Z","signature_status":"signed_v1","first_computed_at":"2026-06-29T01:14:28.826900Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"Learning to Evict from Key-Value Cache","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","headline":"","cross_cats":["cs.LG"],"primary_cat":"cs.CL","authors_text":"Laura Manduchi, Luca Moschella, Ozan Sener","submitted_at":"2026-02-10T19:34:15Z","abstract_excerpt":"The growing size of Large Language Models (LLMs) makes efficient inference challenging, primarily due to the memory demands of the autoregressive Key-Value (KV) cache. Existing eviction or compression methods reduce cost but rely on heuristics, such as recency or past attention scores, which serve only as indirect proxies for a token's future utility and introduce computational overhead. We reframe KV cache eviction as a reinforcement learning (RL) problem: learning to rank tokens by their predicted usefulness for future decoding. To this end, we introduce KV Policy (KVP), a framework of light"},"claims":{"count":0,"items":[],"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"source":{"id":"2602.10238","kind":"arxiv","version":2},"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/2602.10238/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":"2602.10238","created_at":"2026-06-29T01:14:28.826955+00:00"},{"alias_kind":"arxiv_version","alias_value":"2602.10238v2","created_at":"2026-06-29T01:14:28.826955+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2602.10238","created_at":"2026-06-29T01:14:28.826955+00:00"},{"alias_kind":"pith_short_12","alias_value":"VJAH75QSESYA","created_at":"2026-06-29T01:14:28.826955+00:00"},{"alias_kind":"pith_short_16","alias_value":"VJAH75QSESYAYAT5","created_at":"2026-06-29T01:14:28.826955+00:00"},{"alias_kind":"pith_short_8","alias_value":"VJAH75QS","created_at":"2026-06-29T01:14:28.826955+00:00"}],"events":[],"event_summary":{},"paper_claims":[],"inbound_citations":{"count":2,"internal_anchor_count":2,"sample":[{"citing_arxiv_id":"2606.02775","citing_title":"AURA: Action-Gated Memory for Robot Policies at Constant VRAM","ref_index":40,"is_internal_anchor":true},{"citing_arxiv_id":"2605.14037","citing_title":"Self-Pruned Key-Value Attention: Learning When to Write by Predicting Future Utility","ref_index":78,"is_internal_anchor":true}]},"formal_canon":{"evidence_count":0,"sample":[],"anchors":[]},"links":{"html":"https://pith.science/pith/VJAH75QSESYAYAT5N7M3YXPYYD","json":"https://pith.science/pith/VJAH75QSESYAYAT5N7M3YXPYYD.json","graph_json":"https://pith.science/api/pith-number/VJAH75QSESYAYAT5N7M3YXPYYD/graph.json","events_json":"https://pith.science/api/pith-number/VJAH75QSESYAYAT5N7M3YXPYYD/events.json","paper":"https://pith.science/paper/VJAH75QS"},"agent_actions":{"view_html":"https://pith.science/pith/VJAH75QSESYAYAT5N7M3YXPYYD","download_json":"https://pith.science/pith/VJAH75QSESYAYAT5N7M3YXPYYD.json","view_paper":"https://pith.science/paper/VJAH75QS","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2602.10238&json=true","fetch_graph":"https://pith.science/api/pith-number/VJAH75QSESYAYAT5N7M3YXPYYD/graph.json","fetch_events":"https://pith.science/api/pith-number/VJAH75QSESYAYAT5N7M3YXPYYD/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/VJAH75QSESYAYAT5N7M3YXPYYD/action/timestamp_anchor","attest_storage":"https://pith.science/pith/VJAH75QSESYAYAT5N7M3YXPYYD/action/storage_attestation","attest_author":"https://pith.science/pith/VJAH75QSESYAYAT5N7M3YXPYYD/action/author_attestation","sign_citation":"https://pith.science/pith/VJAH75QSESYAYAT5N7M3YXPYYD/action/citation_signature","submit_replication":"https://pith.science/pith/VJAH75QSESYAYAT5N7M3YXPYYD/action/replication_record"}},"created_at":"2026-06-29T01:14:28.826955+00:00","updated_at":"2026-06-29T01:14:28.826955+00:00"}