{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2025:6CZK46H2HMRBX6MJJEAJ55DF2X","short_pith_number":"pith:6CZK46H2","schema_version":"1.0","canonical_sha256":"f0b2ae78fa3b221bf98949009ef465d5fa6e1014c0403c38bfcdc03bf79aba58","source":{"kind":"arxiv","id":"2502.12202","version":2},"attestation_state":"computed","paper":{"title":"To Think or Not to Think: Exploring the Unthinking Vulnerability in Large Reasoning Models","license":"http://creativecommons.org/licenses/by/4.0/","headline":"","cross_cats":["cs.AI","cs.LG"],"primary_cat":"cs.CL","authors_text":"Baoyuan Wu, Hongbao Zhang, Ke Xu, Ruotong Wang, Siwei Lyu, Zihao Zhu","submitted_at":"2025-02-16T10:45:56Z","abstract_excerpt":"Large Reasoning Models (LRMs) are designed to solve complex tasks by generating explicit reasoning traces before producing final answers. However, we reveal a critical vulnerability in LRMs -- termed Unthinking Vulnerability -- wherein the thinking process can be bypassed by manipulating special delimiter tokens. It is empirically demonstrated to be widespread across mainstream LRMs, posing both a significant risk and potential utility, depending on how it is exploited. In this paper, we systematically investigate this vulnerability from both malicious and beneficial perspectives. On the malic"},"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":"2502.12202","kind":"arxiv","version":2},"metadata":{"license":"http://creativecommons.org/licenses/by/4.0/","primary_cat":"cs.CL","submitted_at":"2025-02-16T10:45:56Z","cross_cats_sorted":["cs.AI","cs.LG"],"title_canon_sha256":"3fa588fb9653b7d1350943f800608ce8a1a3e9222badb583f073d6e4d3acbe19","abstract_canon_sha256":"878b98673a6d3923fe88f07fb64fdbdda6c9935ab7d125d067fdf72b367866a7"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-07-05T11:04:21.226824Z","signature_b64":"AC64PzoYdEzS1qkSntChamT7MB5xLUtbqKATcMFIt5BIfufV51rlyJZV5KbMbOFchkOb/08QWrFGIWpKqnRfAA==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"f0b2ae78fa3b221bf98949009ef465d5fa6e1014c0403c38bfcdc03bf79aba58","last_reissued_at":"2026-07-05T11:04:21.226276Z","signature_status":"signed_v1","first_computed_at":"2026-07-05T11:04:21.226276Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"To Think or Not to Think: Exploring the Unthinking Vulnerability in Large Reasoning Models","license":"http://creativecommons.org/licenses/by/4.0/","headline":"","cross_cats":["cs.AI","cs.LG"],"primary_cat":"cs.CL","authors_text":"Baoyuan Wu, Hongbao Zhang, Ke Xu, Ruotong Wang, Siwei Lyu, Zihao Zhu","submitted_at":"2025-02-16T10:45:56Z","abstract_excerpt":"Large Reasoning Models (LRMs) are designed to solve complex tasks by generating explicit reasoning traces before producing final answers. However, we reveal a critical vulnerability in LRMs -- termed Unthinking Vulnerability -- wherein the thinking process can be bypassed by manipulating special delimiter tokens. It is empirically demonstrated to be widespread across mainstream LRMs, posing both a significant risk and potential utility, depending on how it is exploited. In this paper, we systematically investigate this vulnerability from both malicious and beneficial perspectives. On the malic"},"claims":{"count":0,"items":[],"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"source":{"id":"2502.12202","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/2502.12202/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":"2502.12202","created_at":"2026-07-05T11:04:21.226342+00:00"},{"alias_kind":"arxiv_version","alias_value":"2502.12202v2","created_at":"2026-07-05T11:04:21.226342+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2502.12202","created_at":"2026-07-05T11:04:21.226342+00:00"},{"alias_kind":"pith_short_12","alias_value":"6CZK46H2HMRB","created_at":"2026-07-05T11:04:21.226342+00:00"},{"alias_kind":"pith_short_16","alias_value":"6CZK46H2HMRBX6MJ","created_at":"2026-07-05T11:04:21.226342+00:00"},{"alias_kind":"pith_short_8","alias_value":"6CZK46H2","created_at":"2026-07-05T11:04:21.226342+00:00"}],"events":[],"event_summary":{},"paper_claims":[],"inbound_citations":{"count":2,"internal_anchor_count":0,"sample":[{"citing_arxiv_id":"2504.05605","citing_title":"ShadowCoT: Cognitive Hijacking for Stealthy Reasoning Backdoors in LLMs","ref_index":11,"is_internal_anchor":false},{"citing_arxiv_id":"2605.08378","citing_title":"Reinforcement Learning for Scalable and Trustworthy Intelligent Systems","ref_index":206,"is_internal_anchor":false}]},"formal_canon":{"evidence_count":0,"sample":[],"anchors":[]},"links":{"html":"https://pith.science/pith/6CZK46H2HMRBX6MJJEAJ55DF2X","json":"https://pith.science/pith/6CZK46H2HMRBX6MJJEAJ55DF2X.json","graph_json":"https://pith.science/api/pith-number/6CZK46H2HMRBX6MJJEAJ55DF2X/graph.json","events_json":"https://pith.science/api/pith-number/6CZK46H2HMRBX6MJJEAJ55DF2X/events.json","paper":"https://pith.science/paper/6CZK46H2"},"agent_actions":{"view_html":"https://pith.science/pith/6CZK46H2HMRBX6MJJEAJ55DF2X","download_json":"https://pith.science/pith/6CZK46H2HMRBX6MJJEAJ55DF2X.json","view_paper":"https://pith.science/paper/6CZK46H2","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2502.12202&json=true","fetch_graph":"https://pith.science/api/pith-number/6CZK46H2HMRBX6MJJEAJ55DF2X/graph.json","fetch_events":"https://pith.science/api/pith-number/6CZK46H2HMRBX6MJJEAJ55DF2X/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/6CZK46H2HMRBX6MJJEAJ55DF2X/action/timestamp_anchor","attest_storage":"https://pith.science/pith/6CZK46H2HMRBX6MJJEAJ55DF2X/action/storage_attestation","attest_author":"https://pith.science/pith/6CZK46H2HMRBX6MJJEAJ55DF2X/action/author_attestation","sign_citation":"https://pith.science/pith/6CZK46H2HMRBX6MJJEAJ55DF2X/action/citation_signature","submit_replication":"https://pith.science/pith/6CZK46H2HMRBX6MJJEAJ55DF2X/action/replication_record"}},"created_at":"2026-07-05T11:04:21.226342+00:00","updated_at":"2026-07-05T11:04:21.226342+00:00"}