{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2025:V7KDF574ZKYG5HWNKVAEYCYYVG","short_pith_number":"pith:V7KDF574","schema_version":"1.0","canonical_sha256":"afd432f7fccab06e9ecd55404c0b18a99d2849ac06575b302f51650291c0ea18","source":{"kind":"arxiv","id":"2503.00596","version":1},"attestation_state":"computed","paper":{"title":"BadJudge: Backdoor Vulnerabilities of LLM-as-a-Judge","license":"http://creativecommons.org/licenses/by/4.0/","headline":"","cross_cats":["cs.AI","cs.CR"],"primary_cat":"cs.CL","authors_text":"Fei Wang, Muhao Chen, Terry Tong, Zhe Zhao","submitted_at":"2025-03-01T19:35:01Z","abstract_excerpt":"This paper proposes a novel backdoor threat attacking the LLM-as-a-Judge evaluation regime, where the adversary controls both the candidate and evaluator model. The backdoored evaluator victimizes benign users by unfairly assigning inflated scores to adversary. A trivial single token backdoor poisoning 1% of the evaluator training data triples the adversary's score with respect to their legitimate score. We systematically categorize levels of data access corresponding to three real-world settings, (1) web poisoning, (2) malicious annotator, and (3) weight poisoning. These regimes reflect a wea"},"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":"2503.00596","kind":"arxiv","version":1},"metadata":{"license":"http://creativecommons.org/licenses/by/4.0/","primary_cat":"cs.CL","submitted_at":"2025-03-01T19:35:01Z","cross_cats_sorted":["cs.AI","cs.CR"],"title_canon_sha256":"3be07791dfb520344ccf16da82128f9601fd56539518e5a4e4b1356d9f77c85b","abstract_canon_sha256":"77a7899b716cd591ce996eb7a07f820414d7a3265883327325f317bb71a46c1c"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-07-05T10:22:25.812121Z","signature_b64":"Nf2o4/Hbpj9ZGZOrWTZ0cwHIeYksANCQkppj1qVJ6sEB1rWutCqt6ZuQieLfTl7fu/5gn0/7yODhq/a5bA2sDw==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"afd432f7fccab06e9ecd55404c0b18a99d2849ac06575b302f51650291c0ea18","last_reissued_at":"2026-07-05T10:22:25.811625Z","signature_status":"signed_v1","first_computed_at":"2026-07-05T10:22:25.811625Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"BadJudge: Backdoor Vulnerabilities of LLM-as-a-Judge","license":"http://creativecommons.org/licenses/by/4.0/","headline":"","cross_cats":["cs.AI","cs.CR"],"primary_cat":"cs.CL","authors_text":"Fei Wang, Muhao Chen, Terry Tong, Zhe Zhao","submitted_at":"2025-03-01T19:35:01Z","abstract_excerpt":"This paper proposes a novel backdoor threat attacking the LLM-as-a-Judge evaluation regime, where the adversary controls both the candidate and evaluator model. The backdoored evaluator victimizes benign users by unfairly assigning inflated scores to adversary. A trivial single token backdoor poisoning 1% of the evaluator training data triples the adversary's score with respect to their legitimate score. We systematically categorize levels of data access corresponding to three real-world settings, (1) web poisoning, (2) malicious annotator, and (3) weight poisoning. These regimes reflect a wea"},"claims":{"count":0,"items":[],"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"source":{"id":"2503.00596","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/2503.00596/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":"2503.00596","created_at":"2026-07-05T10:22:25.811683+00:00"},{"alias_kind":"arxiv_version","alias_value":"2503.00596v1","created_at":"2026-07-05T10:22:25.811683+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2503.00596","created_at":"2026-07-05T10:22:25.811683+00:00"},{"alias_kind":"pith_short_12","alias_value":"V7KDF574ZKYG","created_at":"2026-07-05T10:22:25.811683+00:00"},{"alias_kind":"pith_short_16","alias_value":"V7KDF574ZKYG5HWN","created_at":"2026-07-05T10:22:25.811683+00:00"},{"alias_kind":"pith_short_8","alias_value":"V7KDF574","created_at":"2026-07-05T10:22:25.811683+00:00"}],"events":[],"event_summary":{},"paper_claims":[],"inbound_citations":{"count":4,"internal_anchor_count":0,"sample":[{"citing_arxiv_id":"2604.06840","citing_title":"MirageBackdoor: A Stealthy Attack that Induces Think-Well-Answer-Wrong Reasoning","ref_index":3,"is_internal_anchor":false},{"citing_arxiv_id":"2604.13602","citing_title":"Reward Hacking in the Era of Large Models: Mechanisms, Emergent Misalignment, Challenges","ref_index":66,"is_internal_anchor":false},{"citing_arxiv_id":"2605.02647","citing_title":"ContextualJailbreak: Evolutionary Red-Teaming via Simulated Conversational Priming","ref_index":32,"is_internal_anchor":false},{"citing_arxiv_id":"2604.23593","citing_title":"When AI reviews science: Can we trust the referee?","ref_index":9,"is_internal_anchor":false}]},"formal_canon":{"evidence_count":0,"sample":[],"anchors":[]},"links":{"html":"https://pith.science/pith/V7KDF574ZKYG5HWNKVAEYCYYVG","json":"https://pith.science/pith/V7KDF574ZKYG5HWNKVAEYCYYVG.json","graph_json":"https://pith.science/api/pith-number/V7KDF574ZKYG5HWNKVAEYCYYVG/graph.json","events_json":"https://pith.science/api/pith-number/V7KDF574ZKYG5HWNKVAEYCYYVG/events.json","paper":"https://pith.science/paper/V7KDF574"},"agent_actions":{"view_html":"https://pith.science/pith/V7KDF574ZKYG5HWNKVAEYCYYVG","download_json":"https://pith.science/pith/V7KDF574ZKYG5HWNKVAEYCYYVG.json","view_paper":"https://pith.science/paper/V7KDF574","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2503.00596&json=true","fetch_graph":"https://pith.science/api/pith-number/V7KDF574ZKYG5HWNKVAEYCYYVG/graph.json","fetch_events":"https://pith.science/api/pith-number/V7KDF574ZKYG5HWNKVAEYCYYVG/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/V7KDF574ZKYG5HWNKVAEYCYYVG/action/timestamp_anchor","attest_storage":"https://pith.science/pith/V7KDF574ZKYG5HWNKVAEYCYYVG/action/storage_attestation","attest_author":"https://pith.science/pith/V7KDF574ZKYG5HWNKVAEYCYYVG/action/author_attestation","sign_citation":"https://pith.science/pith/V7KDF574ZKYG5HWNKVAEYCYYVG/action/citation_signature","submit_replication":"https://pith.science/pith/V7KDF574ZKYG5HWNKVAEYCYYVG/action/replication_record"}},"created_at":"2026-07-05T10:22:25.811683+00:00","updated_at":"2026-07-05T10:22:25.811683+00:00"}