{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2024:MSGLTVDHTDTTGGMBOBEJKUMXPJ","short_pith_number":"pith:MSGLTVDH","schema_version":"1.0","canonical_sha256":"648cb9d46798e733198170489551977a6a08c2cccbda2501f271797324e96162","source":{"kind":"arxiv","id":"2404.05530","version":2},"attestation_state":"computed","paper":{"title":"Best-of-Venom: Attacking RLHF by Injecting Poisoned Preference Data","license":"http://creativecommons.org/licenses/by-sa/4.0/","headline":"","cross_cats":["cs.AI","cs.CR","cs.LG"],"primary_cat":"cs.CL","authors_text":"Dana Alon, Donald Metzler, Tim Baumg\\\"artner, Yang Gao","submitted_at":"2024-04-08T13:59:02Z","abstract_excerpt":"Reinforcement Learning from Human Feedback (RLHF) is a popular method for aligning Language Models (LM) with human values and preferences. RLHF requires a large number of preference pairs as training data, which are often used in both the Supervised Fine-Tuning and Reward Model training and therefore publicly available datasets are commonly used. In this work, we study to what extent a malicious actor can manipulate the LMs generations by poisoning the preferences, i.e., injecting poisonous preference pairs into these datasets and the RLHF training process. We propose strategies to build poiso"},"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":"2404.05530","kind":"arxiv","version":2},"metadata":{"license":"http://creativecommons.org/licenses/by-sa/4.0/","primary_cat":"cs.CL","submitted_at":"2024-04-08T13:59:02Z","cross_cats_sorted":["cs.AI","cs.CR","cs.LG"],"title_canon_sha256":"7b3ec37d0a837b1e99b0ffc0f5b39b99c3c48f98271d4e693047055931ae77ae","abstract_canon_sha256":"5d6519813cdefe53c8391ff14f231a7e8f76885fda52c8aade85a3dd58239ded"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-07-05T08:52:31.529907Z","signature_b64":"EnInu7I+VlAcH0nb8tCWZh8lKuRtMwWOLvNozkyND52afPnFiCPGQszIqFD8wAMyiJzEz75/Db92fQj063esAA==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"648cb9d46798e733198170489551977a6a08c2cccbda2501f271797324e96162","last_reissued_at":"2026-07-05T08:52:31.529504Z","signature_status":"signed_v1","first_computed_at":"2026-07-05T08:52:31.529504Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"Best-of-Venom: Attacking RLHF by Injecting Poisoned Preference Data","license":"http://creativecommons.org/licenses/by-sa/4.0/","headline":"","cross_cats":["cs.AI","cs.CR","cs.LG"],"primary_cat":"cs.CL","authors_text":"Dana Alon, Donald Metzler, Tim Baumg\\\"artner, Yang Gao","submitted_at":"2024-04-08T13:59:02Z","abstract_excerpt":"Reinforcement Learning from Human Feedback (RLHF) is a popular method for aligning Language Models (LM) with human values and preferences. RLHF requires a large number of preference pairs as training data, which are often used in both the Supervised Fine-Tuning and Reward Model training and therefore publicly available datasets are commonly used. In this work, we study to what extent a malicious actor can manipulate the LMs generations by poisoning the preferences, i.e., injecting poisonous preference pairs into these datasets and the RLHF training process. We propose strategies to build poiso"},"claims":{"count":0,"items":[],"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"source":{"id":"2404.05530","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/2404.05530/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":"2404.05530","created_at":"2026-07-05T08:52:31.529558+00:00"},{"alias_kind":"arxiv_version","alias_value":"2404.05530v2","created_at":"2026-07-05T08:52:31.529558+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2404.05530","created_at":"2026-07-05T08:52:31.529558+00:00"},{"alias_kind":"pith_short_12","alias_value":"MSGLTVDHTDTT","created_at":"2026-07-05T08:52:31.529558+00:00"},{"alias_kind":"pith_short_16","alias_value":"MSGLTVDHTDTTGGMB","created_at":"2026-07-05T08:52:31.529558+00:00"},{"alias_kind":"pith_short_8","alias_value":"MSGLTVDH","created_at":"2026-07-05T08:52:31.529558+00:00"}],"events":[],"event_summary":{},"paper_claims":[],"inbound_citations":{"count":6,"internal_anchor_count":0,"sample":[{"citing_arxiv_id":"2606.30666","citing_title":"Reframing AGI Confrontation with Off Earth Autonomy","ref_index":13,"is_internal_anchor":false},{"citing_arxiv_id":"2605.19147","citing_title":"Be Kind, Rewrite: Benign Projections via Rewriting Defend Against LLM Data Poisoning Attacks","ref_index":1,"is_internal_anchor":false},{"citing_arxiv_id":"2507.02850","citing_title":"LLM Hypnosis: Exploiting User Feedback for Unauthorized Knowledge Injection to All Users","ref_index":3,"is_internal_anchor":false},{"citing_arxiv_id":"2603.28281","citing_title":"Corruption-robust Offline Multi-agent Reinforcement Learning From Human Feedback","ref_index":2,"is_internal_anchor":false},{"citing_arxiv_id":"2605.12529","citing_title":"BackFlush: Knowledge-Free Backdoor Detection and Elimination with Watermark Preservation in Large Language Models","ref_index":17,"is_internal_anchor":false},{"citing_arxiv_id":"2605.02495","citing_title":"Efficient Preference Poisoning Attack on Offline RLHF","ref_index":45,"is_internal_anchor":false}]},"formal_canon":{"evidence_count":0,"sample":[],"anchors":[]},"links":{"html":"https://pith.science/pith/MSGLTVDHTDTTGGMBOBEJKUMXPJ","json":"https://pith.science/pith/MSGLTVDHTDTTGGMBOBEJKUMXPJ.json","graph_json":"https://pith.science/api/pith-number/MSGLTVDHTDTTGGMBOBEJKUMXPJ/graph.json","events_json":"https://pith.science/api/pith-number/MSGLTVDHTDTTGGMBOBEJKUMXPJ/events.json","paper":"https://pith.science/paper/MSGLTVDH"},"agent_actions":{"view_html":"https://pith.science/pith/MSGLTVDHTDTTGGMBOBEJKUMXPJ","download_json":"https://pith.science/pith/MSGLTVDHTDTTGGMBOBEJKUMXPJ.json","view_paper":"https://pith.science/paper/MSGLTVDH","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2404.05530&json=true","fetch_graph":"https://pith.science/api/pith-number/MSGLTVDHTDTTGGMBOBEJKUMXPJ/graph.json","fetch_events":"https://pith.science/api/pith-number/MSGLTVDHTDTTGGMBOBEJKUMXPJ/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/MSGLTVDHTDTTGGMBOBEJKUMXPJ/action/timestamp_anchor","attest_storage":"https://pith.science/pith/MSGLTVDHTDTTGGMBOBEJKUMXPJ/action/storage_attestation","attest_author":"https://pith.science/pith/MSGLTVDHTDTTGGMBOBEJKUMXPJ/action/author_attestation","sign_citation":"https://pith.science/pith/MSGLTVDHTDTTGGMBOBEJKUMXPJ/action/citation_signature","submit_replication":"https://pith.science/pith/MSGLTVDHTDTTGGMBOBEJKUMXPJ/action/replication_record"}},"created_at":"2026-07-05T08:52:31.529558+00:00","updated_at":"2026-07-05T08:52:31.529558+00:00"}