{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2025:EUIE5JDLAOFZDRHSUUDJWU5EY5","short_pith_number":"pith:EUIE5JDL","schema_version":"1.0","canonical_sha256":"25104ea46b038b91c4f2a5069b53a4c769c51df6ddd1ff127fc4585c02570d84","source":{"kind":"arxiv","id":"2503.00555","version":2},"attestation_state":"computed","paper":{"title":"Safety Tax: Safety Alignment Makes Your Large Reasoning Models Less Reasonable","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","headline":"","cross_cats":["cs.AI","cs.LG"],"primary_cat":"cs.CR","authors_text":"Fatih Ilhan, Ling Liu, Selim Furkan Tekin, Sihao Hu, Tiansheng Huang, Yichang Xu, Zachary Yahn","submitted_at":"2025-03-01T16:42:01Z","abstract_excerpt":"Safety alignment is an important procedure before the official deployment of a Large Language Model (LLM). While safety alignment has been extensively studied for LLM, there is still a large research gap for Large Reasoning Models (LRMs) that equip with improved reasoning capability. We in this paper systematically examine a simplified pipeline for producing safety aligned LRMs. With our evaluation of various LRMs, we deliver two main findings: i) Safety alignment can be done upon the LRM to restore its safety capability. ii) Safety alignment leads to a degradation of the reasoning capability "},"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":"2503.00555","kind":"arxiv","version":2},"metadata":{"license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","primary_cat":"cs.CR","submitted_at":"2025-03-01T16:42:01Z","cross_cats_sorted":["cs.AI","cs.LG"],"title_canon_sha256":"2d634756b3fb1784a1035670adbc9c49f35bf87f15fb3cc8236f30a9458de23c","abstract_canon_sha256":"ae7a7dc399be054ac39585bc685d23b2f53735924fed6a40da5740edebec3c60"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-07-05T11:16:06.597351Z","signature_b64":"Fiok1UY9Yw6tKeTpeGI+hjBQUKBKuFC0RyZCbBJ4TZ59Pdc2TURPlkW9ECapkrS4YS9NtXBfL+AzQxMp8blqAw==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"25104ea46b038b91c4f2a5069b53a4c769c51df6ddd1ff127fc4585c02570d84","last_reissued_at":"2026-07-05T11:16:06.596788Z","signature_status":"signed_v1","first_computed_at":"2026-07-05T11:16:06.596788Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"Safety Tax: Safety Alignment Makes Your Large Reasoning Models Less Reasonable","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","headline":"","cross_cats":["cs.AI","cs.LG"],"primary_cat":"cs.CR","authors_text":"Fatih Ilhan, Ling Liu, Selim Furkan Tekin, Sihao Hu, Tiansheng Huang, Yichang Xu, Zachary Yahn","submitted_at":"2025-03-01T16:42:01Z","abstract_excerpt":"Safety alignment is an important procedure before the official deployment of a Large Language Model (LLM). While safety alignment has been extensively studied for LLM, there is still a large research gap for Large Reasoning Models (LRMs) that equip with improved reasoning capability. We in this paper systematically examine a simplified pipeline for producing safety aligned LRMs. With our evaluation of various LRMs, we deliver two main findings: i) Safety alignment can be done upon the LRM to restore its safety capability. ii) Safety alignment leads to a degradation of the reasoning capability "},"claims":{"count":0,"items":[],"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"source":{"id":"2503.00555","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/2503.00555/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":"2503.00555","created_at":"2026-07-05T11:16:06.596872+00:00"},{"alias_kind":"arxiv_version","alias_value":"2503.00555v2","created_at":"2026-07-05T11:16:06.596872+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2503.00555","created_at":"2026-07-05T11:16:06.596872+00:00"},{"alias_kind":"pith_short_12","alias_value":"EUIE5JDLAOFZ","created_at":"2026-07-05T11:16:06.596872+00:00"},{"alias_kind":"pith_short_16","alias_value":"EUIE5JDLAOFZDRHS","created_at":"2026-07-05T11:16:06.596872+00:00"},{"alias_kind":"pith_short_8","alias_value":"EUIE5JDL","created_at":"2026-07-05T11:16:06.596872+00:00"}],"events":[],"event_summary":{},"paper_claims":[],"inbound_citations":{"count":27,"internal_anchor_count":0,"sample":[{"citing_arxiv_id":"2606.25442","citing_title":"PolicyAlign: Direct Policy-Based Safety Alignment for Large Language Models","ref_index":20,"is_internal_anchor":false},{"citing_arxiv_id":"2606.25097","citing_title":"Speculative Decoding at Temperature Zero: A Scoped Safety-Invariance Screen with a 48,072-Sample Expansion","ref_index":19,"is_internal_anchor":false},{"citing_arxiv_id":"2606.23075","citing_title":"Safety in Self-Evolving LLM Agent Systems: Threats, Amplification, and Case Studies","ref_index":25,"is_internal_anchor":false},{"citing_arxiv_id":"2606.19755","citing_title":"SafeSpec: Fast and Safe LLM via Dynamic Reflective Sampling","ref_index":5,"is_internal_anchor":false},{"citing_arxiv_id":"2606.11817","citing_title":"Grammar-Constrained Decoding Can Jailbreak LLMs into Generating Malicious Code","ref_index":47,"is_internal_anchor":false},{"citing_arxiv_id":"2606.02530","citing_title":"SafeSteer: Localized On-Policy Distillation for Efficient Safety Alignment","ref_index":6,"is_internal_anchor":false},{"citing_arxiv_id":"2605.30838","citing_title":"COMPASS: Cognitive MCTS-Guided Process Alignment for Safe Search Agents","ref_index":3,"is_internal_anchor":false},{"citing_arxiv_id":"2606.00651","citing_title":"MESA: Improving MoE Safety Alignment via Decentralized Expertise","ref_index":46,"is_internal_anchor":false},{"citing_arxiv_id":"2606.03601","citing_title":"DDOR: Delta Debugging for Explainable Overrefusal Testing and Repair","ref_index":12,"is_internal_anchor":false},{"citing_arxiv_id":"2603.17305","citing_title":"Contrastive Reasoning Alignment: Reinforcement Learning from Hidden Representations","ref_index":12,"is_internal_anchor":false},{"citing_arxiv_id":"2605.20266","citing_title":"A Survey of Large Audio Language Models: Generalization, Trustworthiness, and Outlook","ref_index":177,"is_internal_anchor":false},{"citing_arxiv_id":"2605.21392","citing_title":"VIPER-MCP: Detecting and Exploiting Taint-Style Vulnerabilities in Model Context Protocol Servers","ref_index":41,"is_internal_anchor":false},{"citing_arxiv_id":"2605.16471","citing_title":"From AI-Generated Content to Agentic Action: Security and Safety Threats in Generative AI","ref_index":57,"is_internal_anchor":false},{"citing_arxiv_id":"2605.15239","citing_title":"Reducing the Safety Tax in LLM Safety Alignment with On-Policy Self-Distillation","ref_index":11,"is_internal_anchor":false},{"citing_arxiv_id":"2509.21882","citing_title":"Position: The Hidden Costs and Measurement Gaps of Reinforcement Learning with Verifiable Rewards","ref_index":11,"is_internal_anchor":false},{"citing_arxiv_id":"2510.21285","citing_title":"When Models Outthink Their Safety: Unveiling and Mitigating Self-Jailbreak in Large Reasoning Models","ref_index":6,"is_internal_anchor":false},{"citing_arxiv_id":"2512.00412","citing_title":"Red Teaming Large Reasoning Models","ref_index":1,"is_internal_anchor":false},{"citing_arxiv_id":"2512.05929","citing_title":"LLM Harms: A Taxonomy and Discussion","ref_index":237,"is_internal_anchor":false},{"citing_arxiv_id":"2503.09567","citing_title":"Towards Reasoning Era: A Survey of Long Chain-of-Thought for Reasoning Large Language Models","ref_index":297,"is_internal_anchor":false},{"citing_arxiv_id":"2605.08936","citing_title":"Self-ReSET: Learning to Self-Recover from Unsafe Reasoning Trajectories","ref_index":37,"is_internal_anchor":false},{"citing_arxiv_id":"2605.08930","citing_title":"Internalizing Safety Understanding in Large Reasoning Models via Verification","ref_index":6,"is_internal_anchor":false},{"citing_arxiv_id":"2605.05678","citing_title":"Chain of Risk: Safety Failures in Large Reasoning Models and Mitigation via Adaptive Multi-Principle Steering","ref_index":25,"is_internal_anchor":false},{"citing_arxiv_id":"2604.10990","citing_title":"When Verification Fails: How Compositionally Infeasible Claims Escape Rejection","ref_index":13,"is_internal_anchor":false},{"citing_arxiv_id":"2604.06436","citing_title":"The Defense Trilemma: Why Prompt Injection Defense Wrappers Fail?","ref_index":37,"is_internal_anchor":false},{"citing_arxiv_id":"2604.15725","citing_title":"Reasoning-targeted Jailbreak Attacks on Large Reasoning Models via Semantic Triggers and Psychological Framing","ref_index":30,"is_internal_anchor":false}]},"formal_canon":{"evidence_count":0,"sample":[],"anchors":[]},"links":{"html":"https://pith.science/pith/EUIE5JDLAOFZDRHSUUDJWU5EY5","json":"https://pith.science/pith/EUIE5JDLAOFZDRHSUUDJWU5EY5.json","graph_json":"https://pith.science/api/pith-number/EUIE5JDLAOFZDRHSUUDJWU5EY5/graph.json","events_json":"https://pith.science/api/pith-number/EUIE5JDLAOFZDRHSUUDJWU5EY5/events.json","paper":"https://pith.science/paper/EUIE5JDL"},"agent_actions":{"view_html":"https://pith.science/pith/EUIE5JDLAOFZDRHSUUDJWU5EY5","download_json":"https://pith.science/pith/EUIE5JDLAOFZDRHSUUDJWU5EY5.json","view_paper":"https://pith.science/paper/EUIE5JDL","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2503.00555&json=true","fetch_graph":"https://pith.science/api/pith-number/EUIE5JDLAOFZDRHSUUDJWU5EY5/graph.json","fetch_events":"https://pith.science/api/pith-number/EUIE5JDLAOFZDRHSUUDJWU5EY5/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/EUIE5JDLAOFZDRHSUUDJWU5EY5/action/timestamp_anchor","attest_storage":"https://pith.science/pith/EUIE5JDLAOFZDRHSUUDJWU5EY5/action/storage_attestation","attest_author":"https://pith.science/pith/EUIE5JDLAOFZDRHSUUDJWU5EY5/action/author_attestation","sign_citation":"https://pith.science/pith/EUIE5JDLAOFZDRHSUUDJWU5EY5/action/citation_signature","submit_replication":"https://pith.science/pith/EUIE5JDLAOFZDRHSUUDJWU5EY5/action/replication_record"}},"created_at":"2026-07-05T11:16:06.596872+00:00","updated_at":"2026-07-05T11:16:06.596872+00:00"}