{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2025:CMLU2UTAQ2DJ6XU265IJR3HLCF","short_pith_number":"pith:CMLU2UTA","schema_version":"1.0","canonical_sha256":"13174d526086869f5e9af75098eceb116510e657d9834072881b6d36d2eee454","source":{"kind":"arxiv","id":"2509.26619","version":3},"attestation_state":"computed","paper":{"title":"Searching the Internet for Challenging Benchmarks at Scale","license":"http://creativecommons.org/licenses/by/4.0/","headline":"An epsilon-greedy bandit search over web topics finds the hardest benchmarks after exploring only 6 percent of the space.","cross_cats":["cs.AI"],"primary_cat":"cs.CL","authors_text":"Daniel Deutsch, Mara Finkelstein, Markus Freitag, Parker Riley, Vil\\'em Zouhar, Wenda Xu","submitted_at":"2025-09-30T17:55:47Z","abstract_excerpt":"Many static benchmarks are beginning to saturate: as models rapidly improve, they achieve near-perfect scores on fixed test sets, leaving little headroom to expose genuine model weaknesses -- and even expert-curated challenge sets quickly saturate after hillclimbing. We present a fully automatic framework that searches the Internet at scale to construct challenging benchmarks without human curation. The key insight is to model the Internet as a vast space of topics and formalize the search as a multi-armed bandit problem, where each topic's difficulty is revealed only through expensive sample-"},"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":true},"canonical_record":{"source":{"id":"2509.26619","kind":"arxiv","version":3},"metadata":{"license":"http://creativecommons.org/licenses/by/4.0/","primary_cat":"cs.CL","submitted_at":"2025-09-30T17:55:47Z","cross_cats_sorted":["cs.AI"],"title_canon_sha256":"3eaacefe7ae5dbdc157d21d0b60b7578559504beb12791b68f4a4dc74ee294bf","abstract_canon_sha256":"5fc549e455c477fa9e0de55bb264b961c852ed41497e0edc8e038b015deba7fd"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-05-27T01:04:51.449400Z","signature_b64":"G687P2YjGcr8DWEHw/cuU/iknoZiL40nhgjR+zgOTK394EJz+NqDHQ+0SuM4B7Yr2HRMmJdEGfI2SliMQQKkBQ==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"13174d526086869f5e9af75098eceb116510e657d9834072881b6d36d2eee454","last_reissued_at":"2026-05-27T01:04:51.448585Z","signature_status":"signed_v1","first_computed_at":"2026-05-27T01:04:51.448585Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"Searching the Internet for Challenging Benchmarks at Scale","license":"http://creativecommons.org/licenses/by/4.0/","headline":"An epsilon-greedy bandit search over web topics finds the hardest benchmarks after exploring only 6 percent of the space.","cross_cats":["cs.AI"],"primary_cat":"cs.CL","authors_text":"Daniel Deutsch, Mara Finkelstein, Markus Freitag, Parker Riley, Vil\\'em Zouhar, Wenda Xu","submitted_at":"2025-09-30T17:55:47Z","abstract_excerpt":"Many static benchmarks are beginning to saturate: as models rapidly improve, they achieve near-perfect scores on fixed test sets, leaving little headroom to expose genuine model weaknesses -- and even expert-curated challenge sets quickly saturate after hillclimbing. We present a fully automatic framework that searches the Internet at scale to construct challenging benchmarks without human curation. The key insight is to model the Internet as a vast space of topics and formalize the search as a multi-armed bandit problem, where each topic's difficulty is revealed only through expensive sample-"},"claims":{"count":4,"items":[{"kind":"strongest_claim","text":"Our epsilon-greedy strategy identifies the most challenging topics while exploring only 6% of the search space -- a 100 times cost reduction over exhaustive evaluation.","source":"verdict.strongest_claim","status":"machine_extracted","claim_id":"C1","attestation":"unclaimed"},{"kind":"weakest_assumption","text":"Topic difficulty revealed through sample-and-evaluate queries is robust across independent metrics (GEMBA-SQA and MetricX), languages, and models.","source":"verdict.weakest_assumption","status":"machine_extracted","claim_id":"C2","attestation":"unclaimed"},{"kind":"one_line_summary","text":"An epsilon-greedy multi-armed bandit framework automatically discovers challenging internet topics for benchmarks in machine translation and QA, exploring only 6% of the space for 100x cost reduction.","source":"verdict.one_line_summary","status":"machine_extracted","claim_id":"C3","attestation":"unclaimed"},{"kind":"headline","text":"An epsilon-greedy bandit search over web topics finds the hardest benchmarks after exploring only 6 percent of the space.","source":"verdict.pith_extraction.headline","status":"machine_extracted","claim_id":"C4","attestation":"unclaimed"}],"snapshot_sha256":"26e246d523aab861d310327c3c8fa92ccc31c1b198552413dd891191d174f335"},"source":{"id":"2509.26619","kind":"arxiv","version":3},"verdict":{"id":"3153e4dc-123b-49ee-8fc5-f3cca2f80538","model_set":{"reader":"grok-4.3"},"created_at":"2026-05-18T11:37:45.823519Z","strongest_claim":"Our epsilon-greedy strategy identifies the most challenging topics while exploring only 6% of the search space -- a 100 times cost reduction over exhaustive evaluation.","one_line_summary":"An epsilon-greedy multi-armed bandit framework automatically discovers challenging internet topics for benchmarks in machine translation and QA, exploring only 6% of the space for 100x cost reduction.","pipeline_version":"pith-pipeline@v0.9.0","weakest_assumption":"Topic difficulty revealed through sample-and-evaluate queries is robust across independent metrics (GEMBA-SQA and MetricX), languages, and models.","pith_extraction_headline":"An epsilon-greedy bandit search over web topics finds the hardest benchmarks after exploring only 6 percent of the space."},"integrity":{"clean":true,"summary":{"advisory":0,"critical":0,"by_detector":{},"informational":0},"endpoint":"/pith/2509.26619/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":1,"snapshot_sha256":"af67a604f53c259cee19b61100a6fa4a5cfe13adfa1af17ab010d432ad5aac28"},"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":"2509.26619","created_at":"2026-05-27T01:04:51.448722+00:00"},{"alias_kind":"arxiv_version","alias_value":"2509.26619v3","created_at":"2026-05-27T01:04:51.448722+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2509.26619","created_at":"2026-05-27T01:04:51.448722+00:00"},{"alias_kind":"pith_short_12","alias_value":"CMLU2UTAQ2DJ","created_at":"2026-05-27T01:04:51.448722+00:00"},{"alias_kind":"pith_short_16","alias_value":"CMLU2UTAQ2DJ6XU2","created_at":"2026-05-27T01:04:51.448722+00:00"},{"alias_kind":"pith_short_8","alias_value":"CMLU2UTA","created_at":"2026-05-27T01:04:51.448722+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/CMLU2UTAQ2DJ6XU265IJR3HLCF","json":"https://pith.science/pith/CMLU2UTAQ2DJ6XU265IJR3HLCF.json","graph_json":"https://pith.science/api/pith-number/CMLU2UTAQ2DJ6XU265IJR3HLCF/graph.json","events_json":"https://pith.science/api/pith-number/CMLU2UTAQ2DJ6XU265IJR3HLCF/events.json","paper":"https://pith.science/paper/CMLU2UTA"},"agent_actions":{"view_html":"https://pith.science/pith/CMLU2UTAQ2DJ6XU265IJR3HLCF","download_json":"https://pith.science/pith/CMLU2UTAQ2DJ6XU265IJR3HLCF.json","view_paper":"https://pith.science/paper/CMLU2UTA","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2509.26619&json=true","fetch_graph":"https://pith.science/api/pith-number/CMLU2UTAQ2DJ6XU265IJR3HLCF/graph.json","fetch_events":"https://pith.science/api/pith-number/CMLU2UTAQ2DJ6XU265IJR3HLCF/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/CMLU2UTAQ2DJ6XU265IJR3HLCF/action/timestamp_anchor","attest_storage":"https://pith.science/pith/CMLU2UTAQ2DJ6XU265IJR3HLCF/action/storage_attestation","attest_author":"https://pith.science/pith/CMLU2UTAQ2DJ6XU265IJR3HLCF/action/author_attestation","sign_citation":"https://pith.science/pith/CMLU2UTAQ2DJ6XU265IJR3HLCF/action/citation_signature","submit_replication":"https://pith.science/pith/CMLU2UTAQ2DJ6XU265IJR3HLCF/action/replication_record"}},"created_at":"2026-05-27T01:04:51.448722+00:00","updated_at":"2026-05-27T01:04:51.448722+00:00"}