{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2025:CMH7LXVQICB6IM3SNC7KZOVVZG","short_pith_number":"pith:CMH7LXVQ","schema_version":"1.0","canonical_sha256":"130ff5deb04083e4337268beacbab5c9894f016068cec77710bb4d83238bb6bf","source":{"kind":"arxiv","id":"2507.08794","version":3},"attestation_state":"computed","paper":{"title":"One Token to Fool LLM-as-a-Judge","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","headline":"","cross_cats":["cs.CL"],"primary_cat":"cs.LG","authors_text":"Dian Yu, Dong Yu, Haitao Mi, Haolin Liu, Meijia Chen, Sunyuan Kung, Yulai Zhao","submitted_at":"2025-07-11T17:55:22Z","abstract_excerpt":"Large language models (LLMs) are increasingly trusted as automated judges, assisting evaluation and providing reward signals for training other models, particularly in reference-based settings like Reinforcement Learning with Verifiable Rewards (RLVR). However, we uncover a critical vulnerability even in this reference-based paradigm: generative reward models are systematically susceptible to reward hacking. We find that superficial inputs, which we term ''master keys'' such as non-word symbols (e.g., '':'' or ''.'') or generic reasoning openers (e.g., ''Thought process:'' or ''Let's solve thi"},"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":"2507.08794","kind":"arxiv","version":3},"metadata":{"license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","primary_cat":"cs.LG","submitted_at":"2025-07-11T17:55:22Z","cross_cats_sorted":["cs.CL"],"title_canon_sha256":"24b7c3089a8d4af7855c55ced66a9562ab4d33366eca01afc366a6bf36a2efb4","abstract_canon_sha256":"512365434a1e22c4653849ea95d0f2099bc9304222b712b9350c49ace49d5531"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-06-12T01:08:14.773455Z","signature_b64":"rLx0/nWpF259s0QRfwG+Ks/oX4fYrq7tiYkSUZVpNWQ4Tej9ZfoqxD2S710DjLlN1GHKn4pnRMDMQ9BQWIseDA==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"130ff5deb04083e4337268beacbab5c9894f016068cec77710bb4d83238bb6bf","last_reissued_at":"2026-06-12T01:08:14.772348Z","signature_status":"signed_v1","first_computed_at":"2026-06-12T01:08:14.772348Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"One Token to Fool LLM-as-a-Judge","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","headline":"","cross_cats":["cs.CL"],"primary_cat":"cs.LG","authors_text":"Dian Yu, Dong Yu, Haitao Mi, Haolin Liu, Meijia Chen, Sunyuan Kung, Yulai Zhao","submitted_at":"2025-07-11T17:55:22Z","abstract_excerpt":"Large language models (LLMs) are increasingly trusted as automated judges, assisting evaluation and providing reward signals for training other models, particularly in reference-based settings like Reinforcement Learning with Verifiable Rewards (RLVR). However, we uncover a critical vulnerability even in this reference-based paradigm: generative reward models are systematically susceptible to reward hacking. We find that superficial inputs, which we term ''master keys'' such as non-word symbols (e.g., '':'' or ''.'') or generic reasoning openers (e.g., ''Thought process:'' or ''Let's solve thi"},"claims":{"count":0,"items":[],"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"source":{"id":"2507.08794","kind":"arxiv","version":3},"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/2507.08794/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":"2507.08794","created_at":"2026-06-12T01:08:14.772506+00:00"},{"alias_kind":"arxiv_version","alias_value":"2507.08794v3","created_at":"2026-06-12T01:08:14.772506+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2507.08794","created_at":"2026-06-12T01:08:14.772506+00:00"},{"alias_kind":"pith_short_12","alias_value":"CMH7LXVQICB6","created_at":"2026-06-12T01:08:14.772506+00:00"},{"alias_kind":"pith_short_16","alias_value":"CMH7LXVQICB6IM3S","created_at":"2026-06-12T01:08:14.772506+00:00"},{"alias_kind":"pith_short_8","alias_value":"CMH7LXVQ","created_at":"2026-06-12T01:08:14.772506+00:00"}],"events":[],"event_summary":{},"paper_claims":[],"inbound_citations":{"count":11,"internal_anchor_count":11,"sample":[{"citing_arxiv_id":"2605.29251","citing_title":"Provably Secure Agent Guardrail","ref_index":64,"is_internal_anchor":true},{"citing_arxiv_id":"2510.00915","citing_title":"Reinforcement Learning with Verifiable yet Noisy Rewards under Imperfect Verifiers","ref_index":36,"is_internal_anchor":true},{"citing_arxiv_id":"2411.15594","citing_title":"A Survey on LLM-as-a-Judge","ref_index":218,"is_internal_anchor":true},{"citing_arxiv_id":"2605.10810","citing_title":"Likelihood scoring for continuations of mathematical text: a self-supervised benchmark with tests for shortcut vulnerabilities","ref_index":21,"is_internal_anchor":true},{"citing_arxiv_id":"2605.12667","citing_title":"ODRPO: Ordinal Decompositions of Discrete Rewards for Robust Policy Optimization","ref_index":6,"is_internal_anchor":true},{"citing_arxiv_id":"2605.12667","citing_title":"ODRPO: Ordinal Decompositions of Discrete Rewards for Robust Policy Optimization","ref_index":6,"is_internal_anchor":true},{"citing_arxiv_id":"2604.02686","citing_title":"Beyond Semantic Manipulation: Token-Space Attacks on Reward Models","ref_index":21,"is_internal_anchor":true},{"citing_arxiv_id":"2605.02909","citing_title":"Delay, Plateau, or Collapse: Evaluating the Impact of Systematic Verification Error on RLVR","ref_index":10,"is_internal_anchor":true},{"citing_arxiv_id":"2604.05371","citing_title":"LLM-as-Judge for Semantic Judging of Powerline Segmentation in UAV Inspection","ref_index":14,"is_internal_anchor":true},{"citing_arxiv_id":"2604.18493","citing_title":"Too Correct to Learn: Reinforcement Learning on Saturated Reasoning Data","ref_index":5,"is_internal_anchor":true},{"citing_arxiv_id":"2604.23593","citing_title":"When AI reviews science: Can we trust the referee?","ref_index":95,"is_internal_anchor":true}]},"formal_canon":{"evidence_count":0,"sample":[],"anchors":[]},"links":{"html":"https://pith.science/pith/CMH7LXVQICB6IM3SNC7KZOVVZG","json":"https://pith.science/pith/CMH7LXVQICB6IM3SNC7KZOVVZG.json","graph_json":"https://pith.science/api/pith-number/CMH7LXVQICB6IM3SNC7KZOVVZG/graph.json","events_json":"https://pith.science/api/pith-number/CMH7LXVQICB6IM3SNC7KZOVVZG/events.json","paper":"https://pith.science/paper/CMH7LXVQ"},"agent_actions":{"view_html":"https://pith.science/pith/CMH7LXVQICB6IM3SNC7KZOVVZG","download_json":"https://pith.science/pith/CMH7LXVQICB6IM3SNC7KZOVVZG.json","view_paper":"https://pith.science/paper/CMH7LXVQ","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2507.08794&json=true","fetch_graph":"https://pith.science/api/pith-number/CMH7LXVQICB6IM3SNC7KZOVVZG/graph.json","fetch_events":"https://pith.science/api/pith-number/CMH7LXVQICB6IM3SNC7KZOVVZG/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/CMH7LXVQICB6IM3SNC7KZOVVZG/action/timestamp_anchor","attest_storage":"https://pith.science/pith/CMH7LXVQICB6IM3SNC7KZOVVZG/action/storage_attestation","attest_author":"https://pith.science/pith/CMH7LXVQICB6IM3SNC7KZOVVZG/action/author_attestation","sign_citation":"https://pith.science/pith/CMH7LXVQICB6IM3SNC7KZOVVZG/action/citation_signature","submit_replication":"https://pith.science/pith/CMH7LXVQICB6IM3SNC7KZOVVZG/action/replication_record"}},"created_at":"2026-06-12T01:08:14.772506+00:00","updated_at":"2026-06-12T01:08:14.772506+00:00"}