{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2026:ZLHEVXOVNEZWIOJL6IURSSJHUG","short_pith_number":"pith:ZLHEVXOV","schema_version":"1.0","canonical_sha256":"cace4addd5693364392bf229194927a19821e520bf638167b1a8a22cc80b359b","source":{"kind":"arxiv","id":"2608.03893","version":1},"attestation_state":"computed","paper":{"title":"Cross-Model KV Cache Transfer in LLM Families: A Closed-Form Linear Mapping for Prefill Reuse","license":"http://creativecommons.org/licenses/by/4.0/","headline":"","cross_cats":[],"primary_cat":"cs.LG","authors_text":"Bita Darvish Rouhani, Makesh Tarun Chandran, Maximilian Golub, Mohammad Mahdi Kamani, Pantea Zardoshti, Rasoul Shafipour, Ritchie Zhao, Ritika Borkar, Taekyung Heo","submitted_at":"2026-08-04T16:26:47Z","abstract_excerpt":"Production deployments often swap between different-sized models in a family for cost-quality cascading, mid-conversation switching, and routing, and each swap forces the receiver to repay the prefill from scratch. We propose cross-model KV cache transfer, where the receiver reuses the source's KV cache, skipping prefill. We find that cross-model KV has substantial linear structure across matched-KV pairs, where source and target share KV head count and per-head dimension. On Qwen3 14B->32B, one source layer explains 56% of variance in the target's keys and 32% in values, rising to 79% and 65%"},"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":"2608.03893","kind":"arxiv","version":1},"metadata":{"license":"http://creativecommons.org/licenses/by/4.0/","primary_cat":"cs.LG","submitted_at":"2026-08-04T16:26:47Z","cross_cats_sorted":[],"title_canon_sha256":"61c17fe875afb5a14095a8400eb798db39ee0dce8e98c47931ac71bf99789c61","abstract_canon_sha256":"daf1a8219dac37595d85075176d9328c726fa889563f67e66a88171300c5b933"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-08-05T01:37:20.152824Z","signature_b64":"20xMq0rN8tJ5CNWuLcOcUaDLBEGy/r40M6GuUntNmv7tkX/bAgSBxUarH6LGmd84BfHnybgoNWBQXBGeg49+DQ==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"cace4addd5693364392bf229194927a19821e520bf638167b1a8a22cc80b359b","last_reissued_at":"2026-08-05T01:37:20.151314Z","signature_status":"signed_v1","first_computed_at":"2026-08-05T01:37:20.151314Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"Cross-Model KV Cache Transfer in LLM Families: A Closed-Form Linear Mapping for Prefill Reuse","license":"http://creativecommons.org/licenses/by/4.0/","headline":"","cross_cats":[],"primary_cat":"cs.LG","authors_text":"Bita Darvish Rouhani, Makesh Tarun Chandran, Maximilian Golub, Mohammad Mahdi Kamani, Pantea Zardoshti, Rasoul Shafipour, Ritchie Zhao, Ritika Borkar, Taekyung Heo","submitted_at":"2026-08-04T16:26:47Z","abstract_excerpt":"Production deployments often swap between different-sized models in a family for cost-quality cascading, mid-conversation switching, and routing, and each swap forces the receiver to repay the prefill from scratch. We propose cross-model KV cache transfer, where the receiver reuses the source's KV cache, skipping prefill. We find that cross-model KV has substantial linear structure across matched-KV pairs, where source and target share KV head count and per-head dimension. On Qwen3 14B->32B, one source layer explains 56% of variance in the target's keys and 32% in values, rising to 79% and 65%"},"claims":{"count":0,"items":[],"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"source":{"id":"2608.03893","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/2608.03893/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":"2608.03893","created_at":"2026-08-05T01:37:20.151794+00:00"},{"alias_kind":"arxiv_version","alias_value":"2608.03893v1","created_at":"2026-08-05T01:37:20.151794+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2608.03893","created_at":"2026-08-05T01:37:20.151794+00:00"},{"alias_kind":"pith_short_12","alias_value":"ZLHEVXOVNEZW","created_at":"2026-08-05T01:37:20.151794+00:00"},{"alias_kind":"pith_short_16","alias_value":"ZLHEVXOVNEZWIOJL","created_at":"2026-08-05T01:37:20.151794+00:00"},{"alias_kind":"pith_short_8","alias_value":"ZLHEVXOV","created_at":"2026-08-05T01:37:20.151794+00:00"}],"events":[],"event_summary":{},"paper_claims":[],"inbound_citations":{"count":0,"internal_anchor_count":0,"sample":[]},"formal_canon":{"evidence_count":0,"sample":[],"anchors":[]},"links":{"html":"https://pith.science/pith/ZLHEVXOVNEZWIOJL6IURSSJHUG","json":"https://pith.science/pith/ZLHEVXOVNEZWIOJL6IURSSJHUG.json","graph_json":"https://pith.science/api/pith-number/ZLHEVXOVNEZWIOJL6IURSSJHUG/graph.json","events_json":"https://pith.science/api/pith-number/ZLHEVXOVNEZWIOJL6IURSSJHUG/events.json","paper":"https://pith.science/paper/ZLHEVXOV"},"agent_actions":{"view_html":"https://pith.science/pith/ZLHEVXOVNEZWIOJL6IURSSJHUG","download_json":"https://pith.science/pith/ZLHEVXOVNEZWIOJL6IURSSJHUG.json","view_paper":"https://pith.science/paper/ZLHEVXOV","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2608.03893&json=true","fetch_graph":"https://pith.science/api/pith-number/ZLHEVXOVNEZWIOJL6IURSSJHUG/graph.json","fetch_events":"https://pith.science/api/pith-number/ZLHEVXOVNEZWIOJL6IURSSJHUG/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/ZLHEVXOVNEZWIOJL6IURSSJHUG/action/timestamp_anchor","attest_storage":"https://pith.science/pith/ZLHEVXOVNEZWIOJL6IURSSJHUG/action/storage_attestation","attest_author":"https://pith.science/pith/ZLHEVXOVNEZWIOJL6IURSSJHUG/action/author_attestation","sign_citation":"https://pith.science/pith/ZLHEVXOVNEZWIOJL6IURSSJHUG/action/citation_signature","submit_replication":"https://pith.science/pith/ZLHEVXOVNEZWIOJL6IURSSJHUG/action/replication_record"}},"created_at":"2026-08-05T01:37:20.151794+00:00","updated_at":"2026-08-05T01:37:20.151794+00:00"}