{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2024:6OYI7A3L5PFGPS72FGYFUUNV5R","short_pith_number":"pith:6OYI7A3L","schema_version":"1.0","canonical_sha256":"f3b08f836bebca67cbfa29b05a51b5ec616b0eaa9da7f9bee3e83def67ee7f2c","source":{"kind":"arxiv","id":"2410.09247","version":1},"attestation_state":"computed","paper":{"title":"Benchmark Inflation: Revealing LLM Performance Gaps Using Retro-Holdouts","license":"http://creativecommons.org/licenses/by-sa/4.0/","headline":"","cross_cats":["cs.AI","cs.CL"],"primary_cat":"cs.LG","authors_text":"Cenny Wenner, Clement Neo, Esben Kran, Jacob Haimes, Jason Schreiber, Kunvar Thaman, Vassil Tashev","submitted_at":"2024-10-11T20:46:56Z","abstract_excerpt":"The training data for many Large Language Models (LLMs) is contaminated with test data. This means that public benchmarks used to assess LLMs are compromised, suggesting a performance gap between benchmark scores and actual capabilities. Ideally, a private holdout set could be used to accurately verify scores. Unfortunately, such datasets do not exist for most benchmarks, and post-hoc construction of sufficiently similar datasets is non-trivial. To address these issues, we introduce a systematic methodology for (i) retrospectively constructing a holdout dataset for a target dataset, (ii) demon"},"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":"2410.09247","kind":"arxiv","version":1},"metadata":{"license":"http://creativecommons.org/licenses/by-sa/4.0/","primary_cat":"cs.LG","submitted_at":"2024-10-11T20:46:56Z","cross_cats_sorted":["cs.AI","cs.CL"],"title_canon_sha256":"fde39da6ddb8ad6edd159b5f82ab961732364b12bad98f856610673f0518841d","abstract_canon_sha256":"11218bff621513be954c37ffef7677968450907c7af7b341d4bd343e8fe10a2c"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-07-05T09:20:05.639757Z","signature_b64":"nZGv2IphtdvU4LEwitjVLT4lKZFIbQ8IzeEUCgVP7sjfrH9NVKDyxVnu5qP5Chq2ZrsPY7ptn/pIiGG7D3YcCQ==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"f3b08f836bebca67cbfa29b05a51b5ec616b0eaa9da7f9bee3e83def67ee7f2c","last_reissued_at":"2026-07-05T09:20:05.639362Z","signature_status":"signed_v1","first_computed_at":"2026-07-05T09:20:05.639362Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"Benchmark Inflation: Revealing LLM Performance Gaps Using Retro-Holdouts","license":"http://creativecommons.org/licenses/by-sa/4.0/","headline":"","cross_cats":["cs.AI","cs.CL"],"primary_cat":"cs.LG","authors_text":"Cenny Wenner, Clement Neo, Esben Kran, Jacob Haimes, Jason Schreiber, Kunvar Thaman, Vassil Tashev","submitted_at":"2024-10-11T20:46:56Z","abstract_excerpt":"The training data for many Large Language Models (LLMs) is contaminated with test data. This means that public benchmarks used to assess LLMs are compromised, suggesting a performance gap between benchmark scores and actual capabilities. Ideally, a private holdout set could be used to accurately verify scores. Unfortunately, such datasets do not exist for most benchmarks, and post-hoc construction of sufficiently similar datasets is non-trivial. To address these issues, we introduce a systematic methodology for (i) retrospectively constructing a holdout dataset for a target dataset, (ii) demon"},"claims":{"count":0,"items":[],"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"source":{"id":"2410.09247","kind":"arxiv","version":1},"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/2410.09247/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":"2410.09247","created_at":"2026-07-05T09:20:05.639418+00:00"},{"alias_kind":"arxiv_version","alias_value":"2410.09247v1","created_at":"2026-07-05T09:20:05.639418+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2410.09247","created_at":"2026-07-05T09:20:05.639418+00:00"},{"alias_kind":"pith_short_12","alias_value":"6OYI7A3L5PFG","created_at":"2026-07-05T09:20:05.639418+00:00"},{"alias_kind":"pith_short_16","alias_value":"6OYI7A3L5PFGPS72","created_at":"2026-07-05T09:20:05.639418+00:00"},{"alias_kind":"pith_short_8","alias_value":"6OYI7A3L","created_at":"2026-07-05T09:20:05.639418+00:00"}],"events":[],"event_summary":{},"paper_claims":[],"inbound_citations":{"count":2,"internal_anchor_count":0,"sample":[{"citing_arxiv_id":"2605.14164","citing_title":"Unsteady Metrics and Benchmarking Cultures of AI Model Builders","ref_index":27,"is_internal_anchor":false},{"citing_arxiv_id":"2605.02964","citing_title":"Reward Hacking Benchmark: Measuring Exploits in LLM Agents with Tool Use","ref_index":31,"is_internal_anchor":false}]},"formal_canon":{"evidence_count":0,"sample":[],"anchors":[]},"links":{"html":"https://pith.science/pith/6OYI7A3L5PFGPS72FGYFUUNV5R","json":"https://pith.science/pith/6OYI7A3L5PFGPS72FGYFUUNV5R.json","graph_json":"https://pith.science/api/pith-number/6OYI7A3L5PFGPS72FGYFUUNV5R/graph.json","events_json":"https://pith.science/api/pith-number/6OYI7A3L5PFGPS72FGYFUUNV5R/events.json","paper":"https://pith.science/paper/6OYI7A3L"},"agent_actions":{"view_html":"https://pith.science/pith/6OYI7A3L5PFGPS72FGYFUUNV5R","download_json":"https://pith.science/pith/6OYI7A3L5PFGPS72FGYFUUNV5R.json","view_paper":"https://pith.science/paper/6OYI7A3L","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2410.09247&json=true","fetch_graph":"https://pith.science/api/pith-number/6OYI7A3L5PFGPS72FGYFUUNV5R/graph.json","fetch_events":"https://pith.science/api/pith-number/6OYI7A3L5PFGPS72FGYFUUNV5R/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/6OYI7A3L5PFGPS72FGYFUUNV5R/action/timestamp_anchor","attest_storage":"https://pith.science/pith/6OYI7A3L5PFGPS72FGYFUUNV5R/action/storage_attestation","attest_author":"https://pith.science/pith/6OYI7A3L5PFGPS72FGYFUUNV5R/action/author_attestation","sign_citation":"https://pith.science/pith/6OYI7A3L5PFGPS72FGYFUUNV5R/action/citation_signature","submit_replication":"https://pith.science/pith/6OYI7A3L5PFGPS72FGYFUUNV5R/action/replication_record"}},"created_at":"2026-07-05T09:20:05.639418+00:00","updated_at":"2026-07-05T09:20:05.639418+00:00"}