{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2026:ID2IWZPZ3W7JBARHK3N6AIWTWO","short_pith_number":"pith:ID2IWZPZ","schema_version":"1.0","canonical_sha256":"40f48b65f9ddbe90822756dbe022d3b39c8933b28cca5f190832a9e0ddef8f1c","source":{"kind":"arxiv","id":"2605.30504","version":1},"attestation_state":"computed","paper":{"title":"Auditing LLM Benchmarks with Item Response Theory","license":"http://creativecommons.org/licenses/by-sa/4.0/","headline":"","cross_cats":[],"primary_cat":"cs.CL","authors_text":"Daniel M. Bikel, Sander Land","submitted_at":"2026-05-28T19:38:44Z","abstract_excerpt":"LLM benchmark labels are frozen at release and silently propagated into downstream benchmarks, errors and all. We introduce an Item Response Theory-based indicator that surfaces likely mislabels at 95% precision in the top 200 examples across seven preference and multiple-choice benchmarks using responses from 114 models, outperforming a supervised classifier. We trace these errors to mechanical labeling heuristics, upstream annotation mistakes inherited unchanged from source datasets, and fundamentally ambiguous items without a defensible single label. The same model fit reveals that reward m"},"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":"2605.30504","kind":"arxiv","version":1},"metadata":{"license":"http://creativecommons.org/licenses/by-sa/4.0/","primary_cat":"cs.CL","submitted_at":"2026-05-28T19:38:44Z","cross_cats_sorted":[],"title_canon_sha256":"5f01988124339c1d0df05fba51b06e729b5ecb9a01bef61fe7a742ddf19cff17","abstract_canon_sha256":"f3000f8e55a2d404f69586ced76b2e938783b79e5cfe831c1f00b0bae4dafa4a"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-06-01T01:02:57.968095Z","signature_b64":"Q+W6p4FAwWJBMZRr6t0N9YUEsX4qTEmYBtZ/670xhJXHdE4ijNXUbjNWqW2sT6smTi1DYkCrnVXs3wDdBp5YCA==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"40f48b65f9ddbe90822756dbe022d3b39c8933b28cca5f190832a9e0ddef8f1c","last_reissued_at":"2026-06-01T01:02:57.967168Z","signature_status":"signed_v1","first_computed_at":"2026-06-01T01:02:57.967168Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"Auditing LLM Benchmarks with Item Response Theory","license":"http://creativecommons.org/licenses/by-sa/4.0/","headline":"","cross_cats":[],"primary_cat":"cs.CL","authors_text":"Daniel M. Bikel, Sander Land","submitted_at":"2026-05-28T19:38:44Z","abstract_excerpt":"LLM benchmark labels are frozen at release and silently propagated into downstream benchmarks, errors and all. We introduce an Item Response Theory-based indicator that surfaces likely mislabels at 95% precision in the top 200 examples across seven preference and multiple-choice benchmarks using responses from 114 models, outperforming a supervised classifier. We trace these errors to mechanical labeling heuristics, upstream annotation mistakes inherited unchanged from source datasets, and fundamentally ambiguous items without a defensible single label. The same model fit reveals that reward m"},"claims":{"count":0,"items":[],"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"source":{"id":"2605.30504","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/2605.30504/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":"2605.30504","created_at":"2026-06-01T01:02:57.967338+00:00"},{"alias_kind":"arxiv_version","alias_value":"2605.30504v1","created_at":"2026-06-01T01:02:57.967338+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2605.30504","created_at":"2026-06-01T01:02:57.967338+00:00"},{"alias_kind":"pith_short_12","alias_value":"ID2IWZPZ3W7J","created_at":"2026-06-01T01:02:57.967338+00:00"},{"alias_kind":"pith_short_16","alias_value":"ID2IWZPZ3W7JBARH","created_at":"2026-06-01T01:02:57.967338+00:00"},{"alias_kind":"pith_short_8","alias_value":"ID2IWZPZ","created_at":"2026-06-01T01:02:57.967338+00:00"}],"events":[],"event_summary":{},"paper_claims":[],"inbound_citations":{"count":1,"internal_anchor_count":1,"sample":[{"citing_arxiv_id":"2608.02966","citing_title":"Every Wrong Answer Counts: Option-Level Psychometrics for LLM Multiple-Choice Benchmarks","ref_index":51,"is_internal_anchor":true}]},"formal_canon":{"evidence_count":0,"sample":[],"anchors":[]},"links":{"html":"https://pith.science/pith/ID2IWZPZ3W7JBARHK3N6AIWTWO","json":"https://pith.science/pith/ID2IWZPZ3W7JBARHK3N6AIWTWO.json","graph_json":"https://pith.science/api/pith-number/ID2IWZPZ3W7JBARHK3N6AIWTWO/graph.json","events_json":"https://pith.science/api/pith-number/ID2IWZPZ3W7JBARHK3N6AIWTWO/events.json","paper":"https://pith.science/paper/ID2IWZPZ"},"agent_actions":{"view_html":"https://pith.science/pith/ID2IWZPZ3W7JBARHK3N6AIWTWO","download_json":"https://pith.science/pith/ID2IWZPZ3W7JBARHK3N6AIWTWO.json","view_paper":"https://pith.science/paper/ID2IWZPZ","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2605.30504&json=true","fetch_graph":"https://pith.science/api/pith-number/ID2IWZPZ3W7JBARHK3N6AIWTWO/graph.json","fetch_events":"https://pith.science/api/pith-number/ID2IWZPZ3W7JBARHK3N6AIWTWO/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/ID2IWZPZ3W7JBARHK3N6AIWTWO/action/timestamp_anchor","attest_storage":"https://pith.science/pith/ID2IWZPZ3W7JBARHK3N6AIWTWO/action/storage_attestation","attest_author":"https://pith.science/pith/ID2IWZPZ3W7JBARHK3N6AIWTWO/action/author_attestation","sign_citation":"https://pith.science/pith/ID2IWZPZ3W7JBARHK3N6AIWTWO/action/citation_signature","submit_replication":"https://pith.science/pith/ID2IWZPZ3W7JBARHK3N6AIWTWO/action/replication_record"}},"created_at":"2026-06-01T01:02:57.967338+00:00","updated_at":"2026-06-01T01:02:57.967338+00:00"}