{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2025:AAHYCYP3MKTDEGISAFEGYN3RHQ","short_pith_number":"pith:AAHYCYP3","schema_version":"1.0","canonical_sha256":"000f8161fb62a632191201486c37713c2b3df0d6a48fc7bdfd9c226776c37315","source":{"kind":"arxiv","id":"2502.16797","version":1},"attestation_state":"computed","paper":{"title":"Forecasting Rare Language Model Behaviors","license":"http://creativecommons.org/licenses/by/4.0/","headline":"","cross_cats":[],"primary_cat":"cs.LG","authors_text":"Erik Jones, Ethan Perez, Jan Leike, Jared Kaplan, Jesse Mu, Meg Tong, Mohammed Mahfoud, Mrinank Sharma, Roger Grosse, William Fithian","submitted_at":"2025-02-24T03:16:15Z","abstract_excerpt":"Standard language model evaluations can fail to capture risks that emerge only at deployment scale. For example, a model may produce safe responses during a small-scale beta test, yet reveal dangerous information when processing billions of requests at deployment. To remedy this, we introduce a method to forecast potential risks across orders of magnitude more queries than we test during evaluation. We make forecasts by studying each query's elicitation probability -- the probability the query produces a target behavior -- and demonstrate that the largest observed elicitation probabilities pre"},"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":"2502.16797","kind":"arxiv","version":1},"metadata":{"license":"http://creativecommons.org/licenses/by/4.0/","primary_cat":"cs.LG","submitted_at":"2025-02-24T03:16:15Z","cross_cats_sorted":[],"title_canon_sha256":"a820946f70b2fa6f9a92b9bf98c36081d9af768908ee6b90f6270c151d275fad","abstract_canon_sha256":"6f3f6a0a878587f1c17befe56c6eb5954557778e6605c969a914245e1cf6a861"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-07-05T10:18:59.799614Z","signature_b64":"6KQ8N92uu0nrgoVqAMJFgauKHxzY7LbptQvtrsCqKtLnMIpr5Tejg9guc19MKgFZwmCUfZNcWVoc+ydcd/CfBA==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"000f8161fb62a632191201486c37713c2b3df0d6a48fc7bdfd9c226776c37315","last_reissued_at":"2026-07-05T10:18:59.799123Z","signature_status":"signed_v1","first_computed_at":"2026-07-05T10:18:59.799123Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"Forecasting Rare Language Model Behaviors","license":"http://creativecommons.org/licenses/by/4.0/","headline":"","cross_cats":[],"primary_cat":"cs.LG","authors_text":"Erik Jones, Ethan Perez, Jan Leike, Jared Kaplan, Jesse Mu, Meg Tong, Mohammed Mahfoud, Mrinank Sharma, Roger Grosse, William Fithian","submitted_at":"2025-02-24T03:16:15Z","abstract_excerpt":"Standard language model evaluations can fail to capture risks that emerge only at deployment scale. For example, a model may produce safe responses during a small-scale beta test, yet reveal dangerous information when processing billions of requests at deployment. To remedy this, we introduce a method to forecast potential risks across orders of magnitude more queries than we test during evaluation. We make forecasts by studying each query's elicitation probability -- the probability the query produces a target behavior -- and demonstrate that the largest observed elicitation probabilities pre"},"claims":{"count":0,"items":[],"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"source":{"id":"2502.16797","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/2502.16797/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":"2502.16797","created_at":"2026-07-05T10:18:59.799174+00:00"},{"alias_kind":"arxiv_version","alias_value":"2502.16797v1","created_at":"2026-07-05T10:18:59.799174+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2502.16797","created_at":"2026-07-05T10:18:59.799174+00:00"},{"alias_kind":"pith_short_12","alias_value":"AAHYCYP3MKTD","created_at":"2026-07-05T10:18:59.799174+00:00"},{"alias_kind":"pith_short_16","alias_value":"AAHYCYP3MKTDEGIS","created_at":"2026-07-05T10:18:59.799174+00:00"},{"alias_kind":"pith_short_8","alias_value":"AAHYCYP3","created_at":"2026-07-05T10:18:59.799174+00:00"}],"events":[],"event_summary":{},"paper_claims":[],"inbound_citations":{"count":5,"internal_anchor_count":1,"sample":[{"citing_arxiv_id":"2607.07184","citing_title":"Predicting LLM Safety Before Release by Simulating Deployment","ref_index":17,"is_internal_anchor":true},{"citing_arxiv_id":"2606.11445","citing_title":"Forecasting Future Behavior as a Learning Task","ref_index":144,"is_internal_anchor":false},{"citing_arxiv_id":"2606.00047","citing_title":"Comprehensive AI governance requires addressing non-model gains","ref_index":53,"is_internal_anchor":false},{"citing_arxiv_id":"2606.29623","citing_title":"SCARCE: Scalable Cascade Analysis for Rare-event Characterisation via Embeddings","ref_index":20,"is_internal_anchor":false},{"citing_arxiv_id":"2604.22167","citing_title":"Estimating Tail Risks in Language Model Output Distributions","ref_index":22,"is_internal_anchor":false}]},"formal_canon":{"evidence_count":0,"sample":[],"anchors":[]},"links":{"html":"https://pith.science/pith/AAHYCYP3MKTDEGISAFEGYN3RHQ","json":"https://pith.science/pith/AAHYCYP3MKTDEGISAFEGYN3RHQ.json","graph_json":"https://pith.science/api/pith-number/AAHYCYP3MKTDEGISAFEGYN3RHQ/graph.json","events_json":"https://pith.science/api/pith-number/AAHYCYP3MKTDEGISAFEGYN3RHQ/events.json","paper":"https://pith.science/paper/AAHYCYP3"},"agent_actions":{"view_html":"https://pith.science/pith/AAHYCYP3MKTDEGISAFEGYN3RHQ","download_json":"https://pith.science/pith/AAHYCYP3MKTDEGISAFEGYN3RHQ.json","view_paper":"https://pith.science/paper/AAHYCYP3","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2502.16797&json=true","fetch_graph":"https://pith.science/api/pith-number/AAHYCYP3MKTDEGISAFEGYN3RHQ/graph.json","fetch_events":"https://pith.science/api/pith-number/AAHYCYP3MKTDEGISAFEGYN3RHQ/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/AAHYCYP3MKTDEGISAFEGYN3RHQ/action/timestamp_anchor","attest_storage":"https://pith.science/pith/AAHYCYP3MKTDEGISAFEGYN3RHQ/action/storage_attestation","attest_author":"https://pith.science/pith/AAHYCYP3MKTDEGISAFEGYN3RHQ/action/author_attestation","sign_citation":"https://pith.science/pith/AAHYCYP3MKTDEGISAFEGYN3RHQ/action/citation_signature","submit_replication":"https://pith.science/pith/AAHYCYP3MKTDEGISAFEGYN3RHQ/action/replication_record"}},"created_at":"2026-07-05T10:18:59.799174+00:00","updated_at":"2026-07-05T10:18:59.799174+00:00"}