{"paper":{"title":"EchoKV: Efficient KV Cache Compression via Similarity-Based Reconstruction","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","headline":"EchoKV compresses the KV cache by reconstructing discarded components from retained ones using attention head similarities.","cross_cats":[],"primary_cat":"cs.CL","authors_text":"Qingfu Zhu, Shiyu Ji, Wanxiang Che, Yijun Liu, Yixuan Wang","submitted_at":"2026-03-24T07:58:42Z","abstract_excerpt":"The increasing memory demand of the Key-Value (KV) cache poses a significant bottleneck for Large Language Models (LLMs) in long-context applications. Existing low-rank KV compression methods reduce this footprint by modifying model projections, limiting the flexibility to switch back to standard full-cache inference when sufficient memory is available. In this paper, we propose EchoKV, a flexible KV cache compression framework that supports on-demand transitions from full KV caching to compressed caching. Unlike traditional compression-decompression paradigms, EchoKV utilizes a lightweight ne"},"claims":{"count":4,"items":[{"kind":"strongest_claim","text":"EchoKV consistently outperforms existing methods across multiple compression ratios and backbone models while preserving the throughput of full-cache inference in short-context scenarios.","source":"verdict.strongest_claim","status":"machine_extracted","claim_id":"C1","attestation":"unclaimed"},{"kind":"weakest_assumption","text":"That intrinsic inter-layer and intra-layer similarities among attention heads are sufficiently stable and informative for a lightweight network to accurately reconstruct the discarded KV components without introducing errors that degrade downstream performance.","source":"verdict.weakest_assumption","status":"machine_extracted","claim_id":"C2","attestation":"unclaimed"},{"kind":"one_line_summary","text":"EchoKV compresses LLM KV caches by reconstructing missing components from partial data via inter- and intra-layer attention similarities, outperforming prior methods on LongBench and RULER while supporting on-demand full-cache inference.","source":"verdict.one_line_summary","status":"machine_extracted","claim_id":"C3","attestation":"unclaimed"},{"kind":"headline","text":"EchoKV compresses the KV cache by reconstructing discarded components from retained ones using attention head similarities.","source":"verdict.pith_extraction.headline","status":"machine_extracted","claim_id":"C4","attestation":"unclaimed"}],"snapshot_sha256":"6fda05209f4ea21a0fa3febe88df0a5409a9e2c91b2bdf0973c4c6b145296c55"},"source":{"id":"2603.22910","kind":"arxiv","version":2},"verdict":{"id":"09112ba9-99ed-4b8b-a8e0-b869e68825d5","model_set":{"reader":"grok-4.3"},"created_at":"2026-05-15T01:05:08.778242Z","strongest_claim":"EchoKV consistently outperforms existing methods across multiple compression ratios and backbone models while preserving the throughput of full-cache inference in short-context scenarios.","one_line_summary":"EchoKV compresses LLM KV caches by reconstructing missing components from partial data via inter- and intra-layer attention similarities, outperforming prior methods on LongBench and RULER while supporting on-demand full-cache inference.","pipeline_version":"pith-pipeline@v0.9.0","weakest_assumption":"That intrinsic inter-layer and intra-layer similarities among attention heads are sufficiently stable and informative for a lightweight network to accurately reconstruct the discarded KV components without introducing errors that degrade downstream performance.","pith_extraction_headline":"EchoKV compresses the KV cache by reconstructing discarded components from retained ones using attention head similarities."},"references":{"count":22,"sample":[{"doi":"","year":null,"title":"GQA: Training Generalized Multi-Query Transformer Models from Multi-Head Checkpoints","work_id":"b73ad5b2-e553-4c71-b0c9-67e67ba7b158","ref_index":1,"cited_arxiv_id":"2305.13245","is_internal_anchor":true},{"doi":"","year":null,"title":"xkv: Cross-layer svd for kv-cache compression","work_id":"16dc86cb-7b66-400c-be1f-dd459db6f94e","ref_index":2,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":null,"title":"Palu: Compressing kv-cache with low-rank projection.arXiv preprint arXiv:2407.21118","work_id":"dc6247ce-a2da-431d-a935-3fb13543cb13","ref_index":3,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":null,"title":"Towards Reasoning Era: A Survey of Long Chain-of-Thought for Reasoning Large Language Models","work_id":"0b361fed-cf2a-4b90-b61a-de88de4b8840","ref_index":4,"cited_arxiv_id":"2503.09567","is_internal_anchor":true},{"doi":"","year":null,"title":"Homogeneous keys, heterogeneous values: Exploiting local kv cache asymmetry for long-context llms.arXiv preprint arXiv:2506.05410. Tri Dao","work_id":"4b5ebee2-c5a9-4d4e-a150-51b305167091","ref_index":5,"cited_arxiv_id":"","is_internal_anchor":false}],"resolved_work":22,"snapshot_sha256":"037438b6a30a26a8fe206b2768d274d7802774da5a189c9a422e31b3bd371606","internal_anchors":11},"formal_canon":{"evidence_count":2,"snapshot_sha256":"55a2138617dd87589c4dacb9ca17f9f7c2ba662dcb063e4dc87ac0a0f97a0c17"},"author_claims":{"count":0,"strong_count":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"builder_version":"pith-number-builder-2026-05-17-v1"}