{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2025:MZTARQSUJPBLSD256AT5GA6JLS","short_pith_number":"pith:MZTARQSU","schema_version":"1.0","canonical_sha256":"666608c2544bc2b90f5df027d303c95cbfe469ef7374a3f5cc792a26955f75e1","source":{"kind":"arxiv","id":"2503.18893","version":2},"attestation_state":"computed","paper":{"title":"xKV: Cross-Layer KV-Cache Compression via Aligned Singular Vector Extraction","license":"http://creativecommons.org/licenses/by/4.0/","headline":"","cross_cats":["cs.LG"],"primary_cat":"cs.CL","authors_text":"Chi-Chih Chang, Chien-Yu Lin, Huiqiang Jiang, Hung-Yueh Chiang, Kai-Chiang Wu, Luis Ceze, Mohamed S. Abdelfattah, Wei-Cheng Lin, Xilai Dai, Yash Akhauri, Yucheng Li","submitted_at":"2025-03-24T17:06:37Z","abstract_excerpt":"Long-context Large Language Models (LLMs) enable powerful applications but incur high memory costs due to the key-value states (KV-Cache). Recent studies attempt to share KV-Cache across layers, but these approaches either require expensive pretraining or rely on per-token cross-layer cosine similarity that is often limited in practice. We show, via Centered Kernel Alignment (CKA), that the dominant singular vectors of KV-Cache are well aligned across layers. Motivated by this observation, we propose xKV, a post-training compression method that jointly factorizes grouped-layer KV-Cache into a "},"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":"2503.18893","kind":"arxiv","version":2},"metadata":{"license":"http://creativecommons.org/licenses/by/4.0/","primary_cat":"cs.CL","submitted_at":"2025-03-24T17:06:37Z","cross_cats_sorted":["cs.LG"],"title_canon_sha256":"fb828e6b83211df9c16f52a252a2e3ef944b75a5734843472fd11f3b751a3ccb","abstract_canon_sha256":"bbea4ffde791eee594fcc2148979b7574ca335ceb58d99700dbc1d8e163df910"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-05-28T01:04:27.440866Z","signature_b64":"O6D4Kf8YJNTifyV9NUVJV8xNAflf5WgPPJHKeX9z0I2pJi4xkCHgynWRDBXhJxNXANejs8kg+aOVVUH6aHXEDA==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"666608c2544bc2b90f5df027d303c95cbfe469ef7374a3f5cc792a26955f75e1","last_reissued_at":"2026-05-28T01:04:27.440187Z","signature_status":"signed_v1","first_computed_at":"2026-05-28T01:04:27.440187Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"xKV: Cross-Layer KV-Cache Compression via Aligned Singular Vector Extraction","license":"http://creativecommons.org/licenses/by/4.0/","headline":"","cross_cats":["cs.LG"],"primary_cat":"cs.CL","authors_text":"Chi-Chih Chang, Chien-Yu Lin, Huiqiang Jiang, Hung-Yueh Chiang, Kai-Chiang Wu, Luis Ceze, Mohamed S. Abdelfattah, Wei-Cheng Lin, Xilai Dai, Yash Akhauri, Yucheng Li","submitted_at":"2025-03-24T17:06:37Z","abstract_excerpt":"Long-context Large Language Models (LLMs) enable powerful applications but incur high memory costs due to the key-value states (KV-Cache). Recent studies attempt to share KV-Cache across layers, but these approaches either require expensive pretraining or rely on per-token cross-layer cosine similarity that is often limited in practice. We show, via Centered Kernel Alignment (CKA), that the dominant singular vectors of KV-Cache are well aligned across layers. Motivated by this observation, we propose xKV, a post-training compression method that jointly factorizes grouped-layer KV-Cache into a "},"claims":{"count":0,"items":[],"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"source":{"id":"2503.18893","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/2503.18893/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":"2503.18893","created_at":"2026-05-28T01:04:27.440272+00:00"},{"alias_kind":"arxiv_version","alias_value":"2503.18893v2","created_at":"2026-05-28T01:04:27.440272+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2503.18893","created_at":"2026-05-28T01:04:27.440272+00:00"},{"alias_kind":"pith_short_12","alias_value":"MZTARQSUJPBL","created_at":"2026-05-28T01:04:27.440272+00:00"},{"alias_kind":"pith_short_16","alias_value":"MZTARQSUJPBLSD25","created_at":"2026-05-28T01:04:27.440272+00:00"},{"alias_kind":"pith_short_8","alias_value":"MZTARQSU","created_at":"2026-05-28T01:04:27.440272+00:00"}],"events":[],"event_summary":{},"paper_claims":[],"inbound_citations":{"count":9,"internal_anchor_count":9,"sample":[{"citing_arxiv_id":"2605.12770","citing_title":"WriteSAE: Sparse Autoencoders for Recurrent State","ref_index":8,"is_internal_anchor":true},{"citing_arxiv_id":"2605.16439","citing_title":"KVCapsule: Efficient Sequential KV Cache Compression for Vision-Language Models with Asymmetric Redundancy","ref_index":3,"is_internal_anchor":true},{"citing_arxiv_id":"2605.17757","citing_title":"OSCAR: Offline Spectral Covariance-Aware Rotation for 2-bit KV Cache Quantization","ref_index":44,"is_internal_anchor":true},{"citing_arxiv_id":"2603.22910","citing_title":"EchoKV: Efficient KV Cache Compression via Similarity-Based Reconstruction","ref_index":2,"is_internal_anchor":true},{"citing_arxiv_id":"2604.02525","citing_title":"AdaHOP: Fast and Accurate Low-Precision Training via Outlier-Pattern-Aware Rotation","ref_index":6,"is_internal_anchor":true},{"citing_arxiv_id":"2605.08314","citing_title":"FlashSVD v1.5: Making Low-Rank Transformers Inference Actually Fast","ref_index":22,"is_internal_anchor":true},{"citing_arxiv_id":"2605.10875","citing_title":"Compute Where it Counts: Self Optimizing Language Models","ref_index":3,"is_internal_anchor":true},{"citing_arxiv_id":"2604.21231","citing_title":"SparKV: Overhead-Aware KV Cache Loading for Efficient On-Device LLM Inference","ref_index":52,"is_internal_anchor":true},{"citing_arxiv_id":"2605.02905","citing_title":"eOptShrinkQ: Near-Lossless KV Cache Compression Through Optimal Spectral Denoising and Quantization","ref_index":7,"is_internal_anchor":true}]},"formal_canon":{"evidence_count":0,"sample":[],"anchors":[]},"links":{"html":"https://pith.science/pith/MZTARQSUJPBLSD256AT5GA6JLS","json":"https://pith.science/pith/MZTARQSUJPBLSD256AT5GA6JLS.json","graph_json":"https://pith.science/api/pith-number/MZTARQSUJPBLSD256AT5GA6JLS/graph.json","events_json":"https://pith.science/api/pith-number/MZTARQSUJPBLSD256AT5GA6JLS/events.json","paper":"https://pith.science/paper/MZTARQSU"},"agent_actions":{"view_html":"https://pith.science/pith/MZTARQSUJPBLSD256AT5GA6JLS","download_json":"https://pith.science/pith/MZTARQSUJPBLSD256AT5GA6JLS.json","view_paper":"https://pith.science/paper/MZTARQSU","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2503.18893&json=true","fetch_graph":"https://pith.science/api/pith-number/MZTARQSUJPBLSD256AT5GA6JLS/graph.json","fetch_events":"https://pith.science/api/pith-number/MZTARQSUJPBLSD256AT5GA6JLS/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/MZTARQSUJPBLSD256AT5GA6JLS/action/timestamp_anchor","attest_storage":"https://pith.science/pith/MZTARQSUJPBLSD256AT5GA6JLS/action/storage_attestation","attest_author":"https://pith.science/pith/MZTARQSUJPBLSD256AT5GA6JLS/action/author_attestation","sign_citation":"https://pith.science/pith/MZTARQSUJPBLSD256AT5GA6JLS/action/citation_signature","submit_replication":"https://pith.science/pith/MZTARQSUJPBLSD256AT5GA6JLS/action/replication_record"}},"created_at":"2026-05-28T01:04:27.440272+00:00","updated_at":"2026-05-28T01:04:27.440272+00:00"}