{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2024:UPKZGQLXP4YLYCZXSWYVVY2A74","short_pith_number":"pith:UPKZGQLX","schema_version":"1.0","canonical_sha256":"a3d59341777f30bc0b3795b15ae340ff1a3c8f866c46038c60bbabc2efc1e926","source":{"kind":"arxiv","id":"2403.09636","version":2},"attestation_state":"computed","paper":{"title":"Dynamic Memory Compression: Retrofitting LLMs for Accelerated Inference","license":"http://creativecommons.org/licenses/by/4.0/","headline":"","cross_cats":[],"primary_cat":"cs.CL","authors_text":"Adrian {\\L}a\\'ncucki, David Tarjan, Edoardo M. Ponti, Marcin Chochowski, Piotr Nawrot","submitted_at":"2024-03-14T17:59:26Z","abstract_excerpt":"Transformers have emerged as the backbone of large language models (LLMs). However, generation remains inefficient due to the need to store in memory a cache of key-value representations for past tokens, whose size scales linearly with the input sequence length and batch size. As a solution, we propose Dynamic Memory Compression (DMC), a method for online key-value cache compression at inference time. Most importantly, the model learns to apply different compression ratios in different heads and layers. We retrofit pre-trained LLMs such as Llama 2 (7B, 13B and 70B) into DMC Transformers, achie"},"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":"2403.09636","kind":"arxiv","version":2},"metadata":{"license":"http://creativecommons.org/licenses/by/4.0/","primary_cat":"cs.CL","submitted_at":"2024-03-14T17:59:26Z","cross_cats_sorted":[],"title_canon_sha256":"e56eb19347be8e219fb34002b760bd962e3ad8186676f4b706c09523d769d08f","abstract_canon_sha256":"677f23bf8a97d3bc361e2e6a428354ac02b43fe7580f96543bcf3e3c651f3dd4"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-07-05T08:47:16.653518Z","signature_b64":"Tb4tM+VchLtkCKCvtkWK/K3KQiIrqeTRE+U2MeOyx7TG4IdvQMk7IuNHj6pgAtBGO1KquYOj2nST7ANJ3pfiAw==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"a3d59341777f30bc0b3795b15ae340ff1a3c8f866c46038c60bbabc2efc1e926","last_reissued_at":"2026-07-05T08:47:16.653023Z","signature_status":"signed_v1","first_computed_at":"2026-07-05T08:47:16.653023Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"Dynamic Memory Compression: Retrofitting LLMs for Accelerated Inference","license":"http://creativecommons.org/licenses/by/4.0/","headline":"","cross_cats":[],"primary_cat":"cs.CL","authors_text":"Adrian {\\L}a\\'ncucki, David Tarjan, Edoardo M. Ponti, Marcin Chochowski, Piotr Nawrot","submitted_at":"2024-03-14T17:59:26Z","abstract_excerpt":"Transformers have emerged as the backbone of large language models (LLMs). However, generation remains inefficient due to the need to store in memory a cache of key-value representations for past tokens, whose size scales linearly with the input sequence length and batch size. As a solution, we propose Dynamic Memory Compression (DMC), a method for online key-value cache compression at inference time. Most importantly, the model learns to apply different compression ratios in different heads and layers. We retrofit pre-trained LLMs such as Llama 2 (7B, 13B and 70B) into DMC Transformers, achie"},"claims":{"count":0,"items":[],"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"source":{"id":"2403.09636","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/2403.09636/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":"2403.09636","created_at":"2026-07-05T08:47:16.653088+00:00"},{"alias_kind":"arxiv_version","alias_value":"2403.09636v2","created_at":"2026-07-05T08:47:16.653088+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2403.09636","created_at":"2026-07-05T08:47:16.653088+00:00"},{"alias_kind":"pith_short_12","alias_value":"UPKZGQLXP4YL","created_at":"2026-07-05T08:47:16.653088+00:00"},{"alias_kind":"pith_short_16","alias_value":"UPKZGQLXP4YLYCZX","created_at":"2026-07-05T08:47:16.653088+00:00"},{"alias_kind":"pith_short_8","alias_value":"UPKZGQLX","created_at":"2026-07-05T08:47:16.653088+00:00"}],"events":[],"event_summary":{},"paper_claims":[],"inbound_citations":{"count":8,"internal_anchor_count":1,"sample":[{"citing_arxiv_id":"2607.08032","citing_title":"What to Keep, What to Forget: A Rate--Distortion View of Memory Compaction in LLMs and Agents","ref_index":86,"is_internal_anchor":true},{"citing_arxiv_id":"2606.01563","citing_title":"MomentKV: Closing the Directional Gap in KV Cache Eviction for Long-Context Inference","ref_index":14,"is_internal_anchor":false},{"citing_arxiv_id":"2408.01129","citing_title":"A Survey of Mamba","ref_index":139,"is_internal_anchor":false},{"citing_arxiv_id":"2410.13846","citing_title":"LightTransfer: Your Long-Context LLM is Secretly a Hybrid Model with Effortless Adaptation","ref_index":25,"is_internal_anchor":false},{"citing_arxiv_id":"2506.17310","citing_title":"PaceLLM: Brain-Inspired Large Language Models for Long-Context Understanding","ref_index":21,"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":18,"is_internal_anchor":false},{"citing_arxiv_id":"2605.02905","citing_title":"eOptShrinkQ: Near-Lossless KV Cache Compression Through Optimal Spectral Denoising and Quantization","ref_index":29,"is_internal_anchor":false},{"citing_arxiv_id":"2605.05219","citing_title":"Sparse Prefix Caching for Hybrid and Recurrent LLM Serving","ref_index":22,"is_internal_anchor":false}]},"formal_canon":{"evidence_count":0,"sample":[],"anchors":[]},"links":{"html":"https://pith.science/pith/UPKZGQLXP4YLYCZXSWYVVY2A74","json":"https://pith.science/pith/UPKZGQLXP4YLYCZXSWYVVY2A74.json","graph_json":"https://pith.science/api/pith-number/UPKZGQLXP4YLYCZXSWYVVY2A74/graph.json","events_json":"https://pith.science/api/pith-number/UPKZGQLXP4YLYCZXSWYVVY2A74/events.json","paper":"https://pith.science/paper/UPKZGQLX"},"agent_actions":{"view_html":"https://pith.science/pith/UPKZGQLXP4YLYCZXSWYVVY2A74","download_json":"https://pith.science/pith/UPKZGQLXP4YLYCZXSWYVVY2A74.json","view_paper":"https://pith.science/paper/UPKZGQLX","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2403.09636&json=true","fetch_graph":"https://pith.science/api/pith-number/UPKZGQLXP4YLYCZXSWYVVY2A74/graph.json","fetch_events":"https://pith.science/api/pith-number/UPKZGQLXP4YLYCZXSWYVVY2A74/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/UPKZGQLXP4YLYCZXSWYVVY2A74/action/timestamp_anchor","attest_storage":"https://pith.science/pith/UPKZGQLXP4YLYCZXSWYVVY2A74/action/storage_attestation","attest_author":"https://pith.science/pith/UPKZGQLXP4YLYCZXSWYVVY2A74/action/author_attestation","sign_citation":"https://pith.science/pith/UPKZGQLXP4YLYCZXSWYVVY2A74/action/citation_signature","submit_replication":"https://pith.science/pith/UPKZGQLXP4YLYCZXSWYVVY2A74/action/replication_record"}},"created_at":"2026-07-05T08:47:16.653088+00:00","updated_at":"2026-07-05T08:47:16.653088+00:00"}