{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2025:ITBVVWVFIP5GNKIAOVFBC3MZNM","short_pith_number":"pith:ITBVVWVF","schema_version":"1.0","canonical_sha256":"44c35adaa543fa66a900754a116d996b0683a4f2c5decf49dd8756814aaed401","source":{"kind":"arxiv","id":"2502.12206","version":1},"attestation_state":"computed","paper":{"title":"Evaluating the Paperclip Maximizer: Are RL-Based Language Models More Likely to Pursue Instrumental Goals?","license":"http://creativecommons.org/licenses/by/4.0/","headline":"","cross_cats":["cs.CL","cs.LG"],"primary_cat":"cs.AI","authors_text":"Bryan Hooi, Jiaying Wu, Yuan Sui, Yuexin Li, Yufei He, Yulin Chen","submitted_at":"2025-02-16T16:29:20Z","abstract_excerpt":"As large language models (LLMs) continue to evolve, ensuring their alignment with human goals and values remains a pressing challenge. A key concern is \\textit{instrumental convergence}, where an AI system, in optimizing for a given objective, develops unintended intermediate goals that override the ultimate objective and deviate from human-intended goals. This issue is particularly relevant in reinforcement learning (RL)-trained models, which can generate creative but unintended strategies to maximize rewards. In this paper, we explore instrumental convergence in LLMs by comparing models trai"},"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":"2502.12206","kind":"arxiv","version":1},"metadata":{"license":"http://creativecommons.org/licenses/by/4.0/","primary_cat":"cs.AI","submitted_at":"2025-02-16T16:29:20Z","cross_cats_sorted":["cs.CL","cs.LG"],"title_canon_sha256":"57e46e5977200487f549729b0c81a2e0fcc23893fde017255d31858d91dbbf45","abstract_canon_sha256":"8bbfb903970437f5ff9de515f93460c83b01d84f9aa9fd2d79181eecb0f088ee"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-07-05T10:15:51.917650Z","signature_b64":"JC3zRecpKKLTTl03SpcKckkBRxyU0lHB/H1KNQ08VUZ7gPNYoiL4FAa6hFVlVEk/zXXXD94HLO9jpYt6J1/LDQ==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"44c35adaa543fa66a900754a116d996b0683a4f2c5decf49dd8756814aaed401","last_reissued_at":"2026-07-05T10:15:51.917123Z","signature_status":"signed_v1","first_computed_at":"2026-07-05T10:15:51.917123Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"Evaluating the Paperclip Maximizer: Are RL-Based Language Models More Likely to Pursue Instrumental Goals?","license":"http://creativecommons.org/licenses/by/4.0/","headline":"","cross_cats":["cs.CL","cs.LG"],"primary_cat":"cs.AI","authors_text":"Bryan Hooi, Jiaying Wu, Yuan Sui, Yuexin Li, Yufei He, Yulin Chen","submitted_at":"2025-02-16T16:29:20Z","abstract_excerpt":"As large language models (LLMs) continue to evolve, ensuring their alignment with human goals and values remains a pressing challenge. A key concern is \\textit{instrumental convergence}, where an AI system, in optimizing for a given objective, develops unintended intermediate goals that override the ultimate objective and deviate from human-intended goals. This issue is particularly relevant in reinforcement learning (RL)-trained models, which can generate creative but unintended strategies to maximize rewards. In this paper, we explore instrumental convergence in LLMs by comparing models trai"},"claims":{"count":0,"items":[],"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"source":{"id":"2502.12206","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/2502.12206/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":"2502.12206","created_at":"2026-07-05T10:15:51.917180+00:00"},{"alias_kind":"arxiv_version","alias_value":"2502.12206v1","created_at":"2026-07-05T10:15:51.917180+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2502.12206","created_at":"2026-07-05T10:15:51.917180+00:00"},{"alias_kind":"pith_short_12","alias_value":"ITBVVWVFIP5G","created_at":"2026-07-05T10:15:51.917180+00:00"},{"alias_kind":"pith_short_16","alias_value":"ITBVVWVFIP5GNKIA","created_at":"2026-07-05T10:15:51.917180+00:00"},{"alias_kind":"pith_short_8","alias_value":"ITBVVWVF","created_at":"2026-07-05T10:15:51.917180+00:00"}],"events":[],"event_summary":{},"paper_claims":[],"inbound_citations":{"count":3,"internal_anchor_count":0,"sample":[{"citing_arxiv_id":"2607.01006","citing_title":"Understanding Large Language Models","ref_index":10,"is_internal_anchor":false},{"citing_arxiv_id":"2504.20472","citing_title":"Robustness via Referencing: Defending against Prompt Injection Attacks by Referencing the Executed Instruction","ref_index":14,"is_internal_anchor":false},{"citing_arxiv_id":"2605.06490","citing_title":"Instrumental Choices: Measuring the Propensity of LLM Agents to Pursue Instrumental Behaviors","ref_index":8,"is_internal_anchor":false}]},"formal_canon":{"evidence_count":0,"sample":[],"anchors":[]},"links":{"html":"https://pith.science/pith/ITBVVWVFIP5GNKIAOVFBC3MZNM","json":"https://pith.science/pith/ITBVVWVFIP5GNKIAOVFBC3MZNM.json","graph_json":"https://pith.science/api/pith-number/ITBVVWVFIP5GNKIAOVFBC3MZNM/graph.json","events_json":"https://pith.science/api/pith-number/ITBVVWVFIP5GNKIAOVFBC3MZNM/events.json","paper":"https://pith.science/paper/ITBVVWVF"},"agent_actions":{"view_html":"https://pith.science/pith/ITBVVWVFIP5GNKIAOVFBC3MZNM","download_json":"https://pith.science/pith/ITBVVWVFIP5GNKIAOVFBC3MZNM.json","view_paper":"https://pith.science/paper/ITBVVWVF","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2502.12206&json=true","fetch_graph":"https://pith.science/api/pith-number/ITBVVWVFIP5GNKIAOVFBC3MZNM/graph.json","fetch_events":"https://pith.science/api/pith-number/ITBVVWVFIP5GNKIAOVFBC3MZNM/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/ITBVVWVFIP5GNKIAOVFBC3MZNM/action/timestamp_anchor","attest_storage":"https://pith.science/pith/ITBVVWVFIP5GNKIAOVFBC3MZNM/action/storage_attestation","attest_author":"https://pith.science/pith/ITBVVWVFIP5GNKIAOVFBC3MZNM/action/author_attestation","sign_citation":"https://pith.science/pith/ITBVVWVFIP5GNKIAOVFBC3MZNM/action/citation_signature","submit_replication":"https://pith.science/pith/ITBVVWVFIP5GNKIAOVFBC3MZNM/action/replication_record"}},"created_at":"2026-07-05T10:15:51.917180+00:00","updated_at":"2026-07-05T10:15:51.917180+00:00"}