{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2025:SLN7ING4TPN56ATXAEJOZ2E4MW","short_pith_number":"pith:SLN7ING4","schema_version":"1.0","canonical_sha256":"92dbf434dc9bdbdf02770112ece89c658e63e6a0197552a4bc2a9660807861af","source":{"kind":"arxiv","id":"2502.17599","version":2},"attestation_state":"computed","paper":{"title":"MEDA: Dynamic KV Cache Allocation for Efficient Multimodal Long-Context Inference","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","headline":"","cross_cats":[],"primary_cat":"cs.CL","authors_text":"Che Liu, Hui Shen, Mi Zhang, Xin Wang, Zheda Mai, Zhongwei Wan","submitted_at":"2025-02-24T19:34:52Z","abstract_excerpt":"Long-context Multimodal Large Language Models (MLLMs) that incorporate long text-image and text-video modalities, demand substantial resources as their multimodal Key-Value (KV) caches grow with increasing input lengths, challenging inference efficiency. Existing methods for KV cache compression, in both text-only and multimodal LLMs, have neglected attention density variations across layers, thus often adopting uniform or progressive reduction strategies for layer-wise cache allocation. In this work, we propose MEDA, a dynamic layer-wise KV cache allocation method for efficient multimodal lon"},"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":"2502.17599","kind":"arxiv","version":2},"metadata":{"license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","primary_cat":"cs.CL","submitted_at":"2025-02-24T19:34:52Z","cross_cats_sorted":[],"title_canon_sha256":"c2298a308cd04c5b7e145ccece93f549b1cb5b6f3a64fb1e1718b352bb521b27","abstract_canon_sha256":"9c8127512994b7b6228dacf06d19a40d79d46425aa9e5a87b47cc74cacd2fba3"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-07-05T10:30:31.377829Z","signature_b64":"L0RoImGLm/cSor+Oqs8FOrpOqLPQT0bpc9/FnCQfY2Mi5UnMMsDrkkbb9CyVzqNHwEQ3L0jTBwL1TfqrN5KtAA==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"92dbf434dc9bdbdf02770112ece89c658e63e6a0197552a4bc2a9660807861af","last_reissued_at":"2026-07-05T10:30:31.377302Z","signature_status":"signed_v1","first_computed_at":"2026-07-05T10:30:31.377302Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"MEDA: Dynamic KV Cache Allocation for Efficient Multimodal Long-Context Inference","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","headline":"","cross_cats":[],"primary_cat":"cs.CL","authors_text":"Che Liu, Hui Shen, Mi Zhang, Xin Wang, Zheda Mai, Zhongwei Wan","submitted_at":"2025-02-24T19:34:52Z","abstract_excerpt":"Long-context Multimodal Large Language Models (MLLMs) that incorporate long text-image and text-video modalities, demand substantial resources as their multimodal Key-Value (KV) caches grow with increasing input lengths, challenging inference efficiency. Existing methods for KV cache compression, in both text-only and multimodal LLMs, have neglected attention density variations across layers, thus often adopting uniform or progressive reduction strategies for layer-wise cache allocation. In this work, we propose MEDA, a dynamic layer-wise KV cache allocation method for efficient multimodal lon"},"claims":{"count":0,"items":[],"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"source":{"id":"2502.17599","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/2502.17599/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":"2502.17599","created_at":"2026-07-05T10:30:31.377381+00:00"},{"alias_kind":"arxiv_version","alias_value":"2502.17599v2","created_at":"2026-07-05T10:30:31.377381+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2502.17599","created_at":"2026-07-05T10:30:31.377381+00:00"},{"alias_kind":"pith_short_12","alias_value":"SLN7ING4TPN5","created_at":"2026-07-05T10:30:31.377381+00:00"},{"alias_kind":"pith_short_16","alias_value":"SLN7ING4TPN56ATX","created_at":"2026-07-05T10:30:31.377381+00:00"},{"alias_kind":"pith_short_8","alias_value":"SLN7ING4","created_at":"2026-07-05T10:30:31.377381+00:00"}],"events":[],"event_summary":{},"paper_claims":[],"inbound_citations":{"count":4,"internal_anchor_count":0,"sample":[{"citing_arxiv_id":"2606.08302","citing_title":"HACK++: Towards More Effective Head-Aware Key-Value Compression for Efficient Visual Autoregressive Modeling","ref_index":64,"is_internal_anchor":false},{"citing_arxiv_id":"2605.09649","citing_title":"Make Each Token Count: Towards Improving Long-Context Performance with KV Cache Eviction","ref_index":27,"is_internal_anchor":false},{"citing_arxiv_id":"2604.25642","citing_title":"Prefill-Time Intervention for Mitigating Hallucination in Large Vision-Language Models","ref_index":45,"is_internal_anchor":false},{"citing_arxiv_id":"2605.04075","citing_title":"RetentiveKV: State-Space Memory for Uncertainty-Aware Multimodal KV Cache Eviction","ref_index":50,"is_internal_anchor":false}]},"formal_canon":{"evidence_count":0,"sample":[],"anchors":[]},"links":{"html":"https://pith.science/pith/SLN7ING4TPN56ATXAEJOZ2E4MW","json":"https://pith.science/pith/SLN7ING4TPN56ATXAEJOZ2E4MW.json","graph_json":"https://pith.science/api/pith-number/SLN7ING4TPN56ATXAEJOZ2E4MW/graph.json","events_json":"https://pith.science/api/pith-number/SLN7ING4TPN56ATXAEJOZ2E4MW/events.json","paper":"https://pith.science/paper/SLN7ING4"},"agent_actions":{"view_html":"https://pith.science/pith/SLN7ING4TPN56ATXAEJOZ2E4MW","download_json":"https://pith.science/pith/SLN7ING4TPN56ATXAEJOZ2E4MW.json","view_paper":"https://pith.science/paper/SLN7ING4","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2502.17599&json=true","fetch_graph":"https://pith.science/api/pith-number/SLN7ING4TPN56ATXAEJOZ2E4MW/graph.json","fetch_events":"https://pith.science/api/pith-number/SLN7ING4TPN56ATXAEJOZ2E4MW/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/SLN7ING4TPN56ATXAEJOZ2E4MW/action/timestamp_anchor","attest_storage":"https://pith.science/pith/SLN7ING4TPN56ATXAEJOZ2E4MW/action/storage_attestation","attest_author":"https://pith.science/pith/SLN7ING4TPN56ATXAEJOZ2E4MW/action/author_attestation","sign_citation":"https://pith.science/pith/SLN7ING4TPN56ATXAEJOZ2E4MW/action/citation_signature","submit_replication":"https://pith.science/pith/SLN7ING4TPN56ATXAEJOZ2E4MW/action/replication_record"}},"created_at":"2026-07-05T10:30:31.377381+00:00","updated_at":"2026-07-05T10:30:31.377381+00:00"}