{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2025:3CDPA7ZZGQEM3AYDLEO6ZBHI6J","short_pith_number":"pith:3CDPA7ZZ","schema_version":"1.0","canonical_sha256":"d886f07f393408cd8303591dec84e8f266956bf7f1406d4f75bf3cb861f6aa22","source":{"kind":"arxiv","id":"2506.00668","version":1},"attestation_state":"computed","paper":{"title":"SafeTy Reasoning Elicitation Alignment for Multi-Turn Dialogues","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","headline":"","cross_cats":[],"primary_cat":"cs.CL","authors_text":"Amin Hass, Aolin Ding, Benjamin F Morris, Bhavna Gopal, Changwei Liu, Hai Li, Isaac Jacobson, James Kiessling, Jianyi Zhang, Louis DiValentin, Martin Kuo, Maziyar Baran Pouyan, Nicolas Ramos, Randolph Linderman, Yiran Chen","submitted_at":"2025-05-31T18:38:23Z","abstract_excerpt":"Malicious attackers can exploit large language models (LLMs) by engaging them in multi-turn dialogues to achieve harmful objectives, posing significant safety risks to society. To address this challenge, we propose a novel defense mechanism: SafeTy Reasoning Elicitation Alignment for Multi-Turn Dialogues (STREAM). STREAM defends LLMs against multi-turn attacks while preserving their functional capabilities. Our approach involves constructing a human-annotated dataset, the Safety Reasoning Multi-turn Dialogues dataset, which is used to fine-tune a plug-and-play safety reasoning moderator. This "},"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":"2506.00668","kind":"arxiv","version":1},"metadata":{"license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","primary_cat":"cs.CL","submitted_at":"2025-05-31T18:38:23Z","cross_cats_sorted":[],"title_canon_sha256":"037cbf5dc339102b4004d283326986d0c5b023743894c0de95b147789ed1f02d","abstract_canon_sha256":"360dc3bdea8c84f22f9ba805967ad469ef649d1e407c8723455632849c1172f7"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-07-05T11:13:38.357941Z","signature_b64":"71uSyalmZQDsCIAVUscXhsvMRKlmnkiXXol2p8QHXAP0yRdgVmnzcmW0OozLiYHpJbl+uJggi8IKzPqph8KBDw==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"d886f07f393408cd8303591dec84e8f266956bf7f1406d4f75bf3cb861f6aa22","last_reissued_at":"2026-07-05T11:13:38.357390Z","signature_status":"signed_v1","first_computed_at":"2026-07-05T11:13:38.357390Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"SafeTy Reasoning Elicitation Alignment for Multi-Turn Dialogues","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","headline":"","cross_cats":[],"primary_cat":"cs.CL","authors_text":"Amin Hass, Aolin Ding, Benjamin F Morris, Bhavna Gopal, Changwei Liu, Hai Li, Isaac Jacobson, James Kiessling, Jianyi Zhang, Louis DiValentin, Martin Kuo, Maziyar Baran Pouyan, Nicolas Ramos, Randolph Linderman, Yiran Chen","submitted_at":"2025-05-31T18:38:23Z","abstract_excerpt":"Malicious attackers can exploit large language models (LLMs) by engaging them in multi-turn dialogues to achieve harmful objectives, posing significant safety risks to society. To address this challenge, we propose a novel defense mechanism: SafeTy Reasoning Elicitation Alignment for Multi-Turn Dialogues (STREAM). STREAM defends LLMs against multi-turn attacks while preserving their functional capabilities. Our approach involves constructing a human-annotated dataset, the Safety Reasoning Multi-turn Dialogues dataset, which is used to fine-tune a plug-and-play safety reasoning moderator. This "},"claims":{"count":0,"items":[],"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"source":{"id":"2506.00668","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/2506.00668/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":"2506.00668","created_at":"2026-07-05T11:13:38.357451+00:00"},{"alias_kind":"arxiv_version","alias_value":"2506.00668v1","created_at":"2026-07-05T11:13:38.357451+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2506.00668","created_at":"2026-07-05T11:13:38.357451+00:00"},{"alias_kind":"pith_short_12","alias_value":"3CDPA7ZZGQEM","created_at":"2026-07-05T11:13:38.357451+00:00"},{"alias_kind":"pith_short_16","alias_value":"3CDPA7ZZGQEM3AYD","created_at":"2026-07-05T11:13:38.357451+00:00"},{"alias_kind":"pith_short_8","alias_value":"3CDPA7ZZ","created_at":"2026-07-05T11:13:38.357451+00:00"}],"events":[],"event_summary":{},"paper_claims":[],"inbound_citations":{"count":2,"internal_anchor_count":0,"sample":[{"citing_arxiv_id":"2604.09056","citing_title":"Conversations Risk Detection LLMs in Financial Agents via Multi-Stage Generative Rollout","ref_index":31,"is_internal_anchor":false},{"citing_arxiv_id":"2604.17301","citing_title":"RoTRAG: Rule of Thumb Reasoning for Conversation Harm Detection with Retrieval-Augmented Generation","ref_index":16,"is_internal_anchor":false}]},"formal_canon":{"evidence_count":0,"sample":[],"anchors":[]},"links":{"html":"https://pith.science/pith/3CDPA7ZZGQEM3AYDLEO6ZBHI6J","json":"https://pith.science/pith/3CDPA7ZZGQEM3AYDLEO6ZBHI6J.json","graph_json":"https://pith.science/api/pith-number/3CDPA7ZZGQEM3AYDLEO6ZBHI6J/graph.json","events_json":"https://pith.science/api/pith-number/3CDPA7ZZGQEM3AYDLEO6ZBHI6J/events.json","paper":"https://pith.science/paper/3CDPA7ZZ"},"agent_actions":{"view_html":"https://pith.science/pith/3CDPA7ZZGQEM3AYDLEO6ZBHI6J","download_json":"https://pith.science/pith/3CDPA7ZZGQEM3AYDLEO6ZBHI6J.json","view_paper":"https://pith.science/paper/3CDPA7ZZ","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2506.00668&json=true","fetch_graph":"https://pith.science/api/pith-number/3CDPA7ZZGQEM3AYDLEO6ZBHI6J/graph.json","fetch_events":"https://pith.science/api/pith-number/3CDPA7ZZGQEM3AYDLEO6ZBHI6J/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/3CDPA7ZZGQEM3AYDLEO6ZBHI6J/action/timestamp_anchor","attest_storage":"https://pith.science/pith/3CDPA7ZZGQEM3AYDLEO6ZBHI6J/action/storage_attestation","attest_author":"https://pith.science/pith/3CDPA7ZZGQEM3AYDLEO6ZBHI6J/action/author_attestation","sign_citation":"https://pith.science/pith/3CDPA7ZZGQEM3AYDLEO6ZBHI6J/action/citation_signature","submit_replication":"https://pith.science/pith/3CDPA7ZZGQEM3AYDLEO6ZBHI6J/action/replication_record"}},"created_at":"2026-07-05T11:13:38.357451+00:00","updated_at":"2026-07-05T11:13:38.357451+00:00"}