{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2026:22XURPEFGAH4RGA4TZDH6BOG46","short_pith_number":"pith:22XURPEF","schema_version":"1.0","canonical_sha256":"d6af48bc85300fc8981c9e467f05c6e7b672b3fece6aa6aab0b8fea7b439024b","source":{"kind":"arxiv","id":"2605.14292","version":1},"attestation_state":"computed","paper":{"title":"Minimal-Intervention KV Retention: A Design-Space Study and a Diversity-Penalty Survivor","license":"http://creativecommons.org/licenses/by/4.0/","headline":"A one-function change adding a diversity penalty to the KV retention scorer outperforms seven other compression methods on mathematical reasoning at small budgets.","cross_cats":["cs.CL"],"primary_cat":"cs.LG","authors_text":"Libo Sun, Peixiong He, Po-Wei Harn, Xiao Qin","submitted_at":"2026-05-14T02:50:20Z","abstract_excerpt":"KV-cache compression at small budgets is a crowded design space spanning cache representation, head-wise routing, compression cadence, decoding behavior, and within-budget scoring. We study seven mechanisms across these five families under matched mean cache on long-form mathematical reasoning (MATH-500~\\cite{hendrycks2021math}) with two distilled-reasoning models (Qwen-7B and Llama-8B variants of DeepSeek-R1-Distill~\\cite{deepseek2025r1}) at budgets $b \\in \\{64, 128\\}$. All seven were rejected. We then propose $\\alpha$, a one-function modification to the TriAttention~\\cite{mao2026triattention"},"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":true,"formal_links_present":true},"canonical_record":{"source":{"id":"2605.14292","kind":"arxiv","version":1},"metadata":{"license":"http://creativecommons.org/licenses/by/4.0/","primary_cat":"cs.LG","submitted_at":"2026-05-14T02:50:20Z","cross_cats_sorted":["cs.CL"],"title_canon_sha256":"48db189b73d8a297a185b5d4cdb5c4b40855d810f55f09ca0e1d93235d855602","abstract_canon_sha256":"9d5e25bb07ad90108ba2f3c1e85e157eb494bc4ef58bfac217bf4db9a51bad50"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-05-17T23:39:10.188306Z","signature_b64":"9hQ518Wec+wjOnRfd10qJ2N/Nmx50MDPclQlg+Hi2WttiDkI4BudO/hqiQC+MXYLsorochB+sefpjR7/v1nPCA==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"d6af48bc85300fc8981c9e467f05c6e7b672b3fece6aa6aab0b8fea7b439024b","last_reissued_at":"2026-05-17T23:39:10.187767Z","signature_status":"signed_v1","first_computed_at":"2026-05-17T23:39:10.187767Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"Minimal-Intervention KV Retention: A Design-Space Study and a Diversity-Penalty Survivor","license":"http://creativecommons.org/licenses/by/4.0/","headline":"A one-function change adding a diversity penalty to the KV retention scorer outperforms seven other compression methods on mathematical reasoning at small budgets.","cross_cats":["cs.CL"],"primary_cat":"cs.LG","authors_text":"Libo Sun, Peixiong He, Po-Wei Harn, Xiao Qin","submitted_at":"2026-05-14T02:50:20Z","abstract_excerpt":"KV-cache compression at small budgets is a crowded design space spanning cache representation, head-wise routing, compression cadence, decoding behavior, and within-budget scoring. We study seven mechanisms across these five families under matched mean cache on long-form mathematical reasoning (MATH-500~\\cite{hendrycks2021math}) with two distilled-reasoning models (Qwen-7B and Llama-8B variants of DeepSeek-R1-Distill~\\cite{deepseek2025r1}) at budgets $b \\in \\{64, 128\\}$. All seven were rejected. We then propose $\\alpha$, a one-function modification to the TriAttention~\\cite{mao2026triattention"},"claims":{"count":4,"items":[{"kind":"strongest_claim","text":"With λ = 0.5, α clears Bonferroni on two of the four (model, budget) cells (Qwen b=128 and Llama b=64), no cell is significantly negative, and the pre-registered Branch A triggers.","source":"verdict.strongest_claim","status":"machine_extracted","claim_id":"C1","attestation":"unclaimed"},{"kind":"weakest_assumption","text":"That the pre-registered tuning on the development split and confirmation on the held-out split of MATH-500 with these two specific models generalizes beyond the tested regime and tasks.","source":"verdict.weakest_assumption","status":"machine_extracted","claim_id":"C2","attestation":"unclaimed"},{"kind":"one_line_summary","text":"A one-function modification to the TriAttention retention scorer using greedy selection under a V-space redundancy penalty outperforms seven matched mechanisms on long-form math reasoning at budgets 64 and 128.","source":"verdict.one_line_summary","status":"machine_extracted","claim_id":"C3","attestation":"unclaimed"},{"kind":"headline","text":"A one-function change adding a diversity penalty to the KV retention scorer outperforms seven other compression methods on mathematical reasoning at small budgets.","source":"verdict.pith_extraction.headline","status":"machine_extracted","claim_id":"C4","attestation":"unclaimed"}],"snapshot_sha256":"36dec26da357fcad31319f295ee60911215f9ba38c9dca9d8204c9e8afa480e9"},"source":{"id":"2605.14292","kind":"arxiv","version":1},"verdict":{"id":"c0723674-7732-4c7a-95ea-f6f66c405430","model_set":{"reader":"grok-4.3"},"created_at":"2026-05-15T02:14:00.797973Z","strongest_claim":"With λ = 0.5, α clears Bonferroni on two of the four (model, budget) cells (Qwen b=128 and Llama b=64), no cell is significantly negative, and the pre-registered Branch A triggers.","one_line_summary":"A one-function modification to the TriAttention retention scorer using greedy selection under a V-space redundancy penalty outperforms seven matched mechanisms on long-form math reasoning at budgets 64 and 128.","pipeline_version":"pith-pipeline@v0.9.0","weakest_assumption":"That the pre-registered tuning on the development split and confirmation on the held-out split of MATH-500 with these two specific models generalizes beyond the tested regime and tasks.","pith_extraction_headline":"A one-function change adding a diversity penalty to the KV retention scorer outperforms seven other compression methods on mathematical reasoning at small budgets."},"references":{"count":28,"sample":[{"doi":"","year":2024,"title":"LongBench: A bilingual, multitask benchmark for long context understanding","work_id":"5f38cf20-6ae2-4609-92f0-e8288b3a50e4","ref_index":1,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2021,"title":"Accounting for variance in machine learning benchmarks","work_id":"d82b95de-3307-4be8-abd8-9eb022ff5682","ref_index":2,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2024,"title":"Kelly, J. R. Reducing transformer key-value cache size with cross-layer attention, 2024","work_id":"22fddcfe-1474-4671-874e-7ef0328c416e","ref_index":3,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2024,"title":"PyramidKV: Dynamic KV cache compression based on pyramidal information funneling, 2024","work_id":"4439c820-ec96-427f-9d23-6769184dddef","ref_index":4,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2025,"title":"R-KV: Redundancy-aware KV cache compression for reasoning models","work_id":"881ebad1-d4d2-4bf4-ad80-1325f19526a8","ref_index":5,"cited_arxiv_id":"","is_internal_anchor":false}],"resolved_work":28,"snapshot_sha256":"6a96c0f72a5cf01ba510244c3b4f52f2158f62e68fde578fc66ab23292ebc30e","internal_anchors":0},"formal_canon":{"evidence_count":1,"snapshot_sha256":"d0018cf4532f107d63e42ac11e62e763a4d272f5a1e86392c8a1a1e0f12ef5c0"},"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":"2605.14292","created_at":"2026-05-17T23:39:10.187872+00:00"},{"alias_kind":"arxiv_version","alias_value":"2605.14292v1","created_at":"2026-05-17T23:39:10.187872+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2605.14292","created_at":"2026-05-17T23:39:10.187872+00:00"},{"alias_kind":"pith_short_12","alias_value":"22XURPEFGAH4","created_at":"2026-05-18T12:33:37.589309+00:00"},{"alias_kind":"pith_short_16","alias_value":"22XURPEFGAH4RGA4","created_at":"2026-05-18T12:33:37.589309+00:00"},{"alias_kind":"pith_short_8","alias_value":"22XURPEF","created_at":"2026-05-18T12:33:37.589309+00:00"}],"events":[],"event_summary":{},"paper_claims":[],"inbound_citations":{"count":0,"internal_anchor_count":0,"sample":[]},"formal_canon":{"evidence_count":1,"sample":[],"anchors":[]},"links":{"html":"https://pith.science/pith/22XURPEFGAH4RGA4TZDH6BOG46","json":"https://pith.science/pith/22XURPEFGAH4RGA4TZDH6BOG46.json","graph_json":"https://pith.science/api/pith-number/22XURPEFGAH4RGA4TZDH6BOG46/graph.json","events_json":"https://pith.science/api/pith-number/22XURPEFGAH4RGA4TZDH6BOG46/events.json","paper":"https://pith.science/paper/22XURPEF"},"agent_actions":{"view_html":"https://pith.science/pith/22XURPEFGAH4RGA4TZDH6BOG46","download_json":"https://pith.science/pith/22XURPEFGAH4RGA4TZDH6BOG46.json","view_paper":"https://pith.science/paper/22XURPEF","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2605.14292&json=true","fetch_graph":"https://pith.science/api/pith-number/22XURPEFGAH4RGA4TZDH6BOG46/graph.json","fetch_events":"https://pith.science/api/pith-number/22XURPEFGAH4RGA4TZDH6BOG46/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/22XURPEFGAH4RGA4TZDH6BOG46/action/timestamp_anchor","attest_storage":"https://pith.science/pith/22XURPEFGAH4RGA4TZDH6BOG46/action/storage_attestation","attest_author":"https://pith.science/pith/22XURPEFGAH4RGA4TZDH6BOG46/action/author_attestation","sign_citation":"https://pith.science/pith/22XURPEFGAH4RGA4TZDH6BOG46/action/citation_signature","submit_replication":"https://pith.science/pith/22XURPEFGAH4RGA4TZDH6BOG46/action/replication_record"}},"created_at":"2026-05-17T23:39:10.187872+00:00","updated_at":"2026-05-17T23:39:10.187872+00:00"}