{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2025:TXE5RINRCSR2ATRENAQKEH6XJZ","short_pith_number":"pith:TXE5RINR","schema_version":"1.0","canonical_sha256":"9dc9d8a1b114a3a04e246820a21fd74e7b8c033daca0c1d2929da24e3d6ad911","source":{"kind":"arxiv","id":"2509.26600","version":2},"attestation_state":"computed","paper":{"title":"When LLMs Benchmark Themselves: Deconstructing Self-Bias in Automated Evaluation","license":"http://creativecommons.org/licenses/by/4.0/","headline":"","cross_cats":["cs.AI"],"primary_cat":"cs.CL","authors_text":"Daniel Deutsch, Markus Freitag, Sweta Agrawal, Vil\\'em Zouhar, Wenda Xu","submitted_at":"2025-09-30T17:48:35Z","abstract_excerpt":"As LLMs rapidly saturate existing benchmarks, automated benchmark creation using LLMs (LLM-as-a-benchmark) -- where a model generates test inputs (LLM-as-a-testset) and evaluates outputs (LLM-as-an-evaluator) -- has gained traction as a cheap alternative to human curation. We show that this paradigm has a fundamental problem: LLM-generated benchmarks systematically favor the model that created them. Using machine translation as our primary testbed, we find that self-bias arises from two additive sources, LLM-as-a-testset and LLM-as-an-evaluator, and their combination amplifies the effect. Cruc"},"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":"2509.26600","kind":"arxiv","version":2},"metadata":{"license":"http://creativecommons.org/licenses/by/4.0/","primary_cat":"cs.CL","submitted_at":"2025-09-30T17:48:35Z","cross_cats_sorted":["cs.AI"],"title_canon_sha256":"1e59ca5db1d5817ff0e699fcd987382506aaaeaf1d77b7b02601f2c0a7a45953","abstract_canon_sha256":"0414a704ea17279559225647bea6190ac0ee3aed519286199bad6f87825fbabf"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-05-27T01:04:51.382963Z","signature_b64":"uLOHw1wQDU2f2/EejIH/AOdTr3J7Vaq0T+1H5WUbO/XujGclenSY0fnHaPcR9/k3nAiN9MhmyueBGQ3hXTuiAA==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"9dc9d8a1b114a3a04e246820a21fd74e7b8c033daca0c1d2929da24e3d6ad911","last_reissued_at":"2026-05-27T01:04:51.382102Z","signature_status":"signed_v1","first_computed_at":"2026-05-27T01:04:51.382102Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"When LLMs Benchmark Themselves: Deconstructing Self-Bias in Automated Evaluation","license":"http://creativecommons.org/licenses/by/4.0/","headline":"","cross_cats":["cs.AI"],"primary_cat":"cs.CL","authors_text":"Daniel Deutsch, Markus Freitag, Sweta Agrawal, Vil\\'em Zouhar, Wenda Xu","submitted_at":"2025-09-30T17:48:35Z","abstract_excerpt":"As LLMs rapidly saturate existing benchmarks, automated benchmark creation using LLMs (LLM-as-a-benchmark) -- where a model generates test inputs (LLM-as-a-testset) and evaluates outputs (LLM-as-an-evaluator) -- has gained traction as a cheap alternative to human curation. We show that this paradigm has a fundamental problem: LLM-generated benchmarks systematically favor the model that created them. Using machine translation as our primary testbed, we find that self-bias arises from two additive sources, LLM-as-a-testset and LLM-as-an-evaluator, and their combination amplifies the effect. Cruc"},"claims":{"count":0,"items":[],"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"source":{"id":"2509.26600","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/2509.26600/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":"2509.26600","created_at":"2026-05-27T01:04:51.382222+00:00"},{"alias_kind":"arxiv_version","alias_value":"2509.26600v2","created_at":"2026-05-27T01:04:51.382222+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2509.26600","created_at":"2026-05-27T01:04:51.382222+00:00"},{"alias_kind":"pith_short_12","alias_value":"TXE5RINRCSR2","created_at":"2026-05-27T01:04:51.382222+00:00"},{"alias_kind":"pith_short_16","alias_value":"TXE5RINRCSR2ATRE","created_at":"2026-05-27T01:04:51.382222+00:00"},{"alias_kind":"pith_short_8","alias_value":"TXE5RINR","created_at":"2026-05-27T01:04:51.382222+00:00"}],"events":[],"event_summary":{},"paper_claims":[],"inbound_citations":{"count":1,"internal_anchor_count":1,"sample":[{"citing_arxiv_id":"2605.22785","citing_title":"Evaluating Commercial AI Chatbots as News Intermediaries","ref_index":64,"is_internal_anchor":true}]},"formal_canon":{"evidence_count":0,"sample":[],"anchors":[]},"links":{"html":"https://pith.science/pith/TXE5RINRCSR2ATRENAQKEH6XJZ","json":"https://pith.science/pith/TXE5RINRCSR2ATRENAQKEH6XJZ.json","graph_json":"https://pith.science/api/pith-number/TXE5RINRCSR2ATRENAQKEH6XJZ/graph.json","events_json":"https://pith.science/api/pith-number/TXE5RINRCSR2ATRENAQKEH6XJZ/events.json","paper":"https://pith.science/paper/TXE5RINR"},"agent_actions":{"view_html":"https://pith.science/pith/TXE5RINRCSR2ATRENAQKEH6XJZ","download_json":"https://pith.science/pith/TXE5RINRCSR2ATRENAQKEH6XJZ.json","view_paper":"https://pith.science/paper/TXE5RINR","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2509.26600&json=true","fetch_graph":"https://pith.science/api/pith-number/TXE5RINRCSR2ATRENAQKEH6XJZ/graph.json","fetch_events":"https://pith.science/api/pith-number/TXE5RINRCSR2ATRENAQKEH6XJZ/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/TXE5RINRCSR2ATRENAQKEH6XJZ/action/timestamp_anchor","attest_storage":"https://pith.science/pith/TXE5RINRCSR2ATRENAQKEH6XJZ/action/storage_attestation","attest_author":"https://pith.science/pith/TXE5RINRCSR2ATRENAQKEH6XJZ/action/author_attestation","sign_citation":"https://pith.science/pith/TXE5RINRCSR2ATRENAQKEH6XJZ/action/citation_signature","submit_replication":"https://pith.science/pith/TXE5RINRCSR2ATRENAQKEH6XJZ/action/replication_record"}},"created_at":"2026-05-27T01:04:51.382222+00:00","updated_at":"2026-05-27T01:04:51.382222+00:00"}