{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2025:CIJF5WPABF5Q235VXL4RVCS554","short_pith_number":"pith:CIJF5WPA","schema_version":"1.0","canonical_sha256":"12125ed9e0097b0d6fb5baf91a8a5def0c91fd0b8d189dc6f9a37178ef8d717f","source":{"kind":"arxiv","id":"2504.03185","version":1},"attestation_state":"computed","paper":{"title":"Learning Natural Language Constraints for Safe Reinforcement Learning of Language Agents","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","headline":"","cross_cats":["cs.AI"],"primary_cat":"cs.CL","authors_text":"Chen Wang, Jaymari Chua, Lina Yao","submitted_at":"2025-04-04T05:26:28Z","abstract_excerpt":"Generalizable alignment is a core challenge for deploying Large Language Models (LLMs) safely in real-world NLP applications. Current alignment methods, including Reinforcement Learning from Human Feedback (RLHF), often fail to guarantee constraint satisfaction outside their training distribution due to their reliance on implicit, post-hoc preferences. Inspired by a paradigm shift to first curate data before tuning, we introduce a new framework for safe language alignment that learns natural language constraints from positive and negative demonstrations as a primary step. From inferring both a"},"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":"2504.03185","kind":"arxiv","version":1},"metadata":{"license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","primary_cat":"cs.CL","submitted_at":"2025-04-04T05:26:28Z","cross_cats_sorted":["cs.AI"],"title_canon_sha256":"fc7b562a11fe604196ec625a782c5a7bffae43874359578ec1b2be2ec06e5ff8","abstract_canon_sha256":"e18eaed317c8c0f424e02ed782c8aec673505e41a96cab742f8f2877af4b2835"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-07-05T10:44:27.722485Z","signature_b64":"tco2YWQoyJV1MD3/0hYTzoV98WsPp3IIOGX1WR0kSr+3TJr0TYEWDGcWFyMF6/Idd1eVR3b+lki3aBwQTvb3Dg==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"12125ed9e0097b0d6fb5baf91a8a5def0c91fd0b8d189dc6f9a37178ef8d717f","last_reissued_at":"2026-07-05T10:44:27.722020Z","signature_status":"signed_v1","first_computed_at":"2026-07-05T10:44:27.722020Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"Learning Natural Language Constraints for Safe Reinforcement Learning of Language Agents","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","headline":"","cross_cats":["cs.AI"],"primary_cat":"cs.CL","authors_text":"Chen Wang, Jaymari Chua, Lina Yao","submitted_at":"2025-04-04T05:26:28Z","abstract_excerpt":"Generalizable alignment is a core challenge for deploying Large Language Models (LLMs) safely in real-world NLP applications. Current alignment methods, including Reinforcement Learning from Human Feedback (RLHF), often fail to guarantee constraint satisfaction outside their training distribution due to their reliance on implicit, post-hoc preferences. Inspired by a paradigm shift to first curate data before tuning, we introduce a new framework for safe language alignment that learns natural language constraints from positive and negative demonstrations as a primary step. From inferring both a"},"claims":{"count":0,"items":[],"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"source":{"id":"2504.03185","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/2504.03185/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":"2504.03185","created_at":"2026-07-05T10:44:27.722083+00:00"},{"alias_kind":"arxiv_version","alias_value":"2504.03185v1","created_at":"2026-07-05T10:44:27.722083+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2504.03185","created_at":"2026-07-05T10:44:27.722083+00:00"},{"alias_kind":"pith_short_12","alias_value":"CIJF5WPABF5Q","created_at":"2026-07-05T10:44:27.722083+00:00"},{"alias_kind":"pith_short_16","alias_value":"CIJF5WPABF5Q235V","created_at":"2026-07-05T10:44:27.722083+00:00"},{"alias_kind":"pith_short_8","alias_value":"CIJF5WPA","created_at":"2026-07-05T10:44:27.722083+00:00"}],"events":[],"event_summary":{},"paper_claims":[],"inbound_citations":{"count":1,"internal_anchor_count":1,"sample":[{"citing_arxiv_id":"2504.14520","citing_title":"Meta-Thinking in LLMs via Multi-Agent Reinforcement Learning: A Survey","ref_index":42,"is_internal_anchor":true}]},"formal_canon":{"evidence_count":0,"sample":[],"anchors":[]},"links":{"html":"https://pith.science/pith/CIJF5WPABF5Q235VXL4RVCS554","json":"https://pith.science/pith/CIJF5WPABF5Q235VXL4RVCS554.json","graph_json":"https://pith.science/api/pith-number/CIJF5WPABF5Q235VXL4RVCS554/graph.json","events_json":"https://pith.science/api/pith-number/CIJF5WPABF5Q235VXL4RVCS554/events.json","paper":"https://pith.science/paper/CIJF5WPA"},"agent_actions":{"view_html":"https://pith.science/pith/CIJF5WPABF5Q235VXL4RVCS554","download_json":"https://pith.science/pith/CIJF5WPABF5Q235VXL4RVCS554.json","view_paper":"https://pith.science/paper/CIJF5WPA","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2504.03185&json=true","fetch_graph":"https://pith.science/api/pith-number/CIJF5WPABF5Q235VXL4RVCS554/graph.json","fetch_events":"https://pith.science/api/pith-number/CIJF5WPABF5Q235VXL4RVCS554/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/CIJF5WPABF5Q235VXL4RVCS554/action/timestamp_anchor","attest_storage":"https://pith.science/pith/CIJF5WPABF5Q235VXL4RVCS554/action/storage_attestation","attest_author":"https://pith.science/pith/CIJF5WPABF5Q235VXL4RVCS554/action/author_attestation","sign_citation":"https://pith.science/pith/CIJF5WPABF5Q235VXL4RVCS554/action/citation_signature","submit_replication":"https://pith.science/pith/CIJF5WPABF5Q235VXL4RVCS554/action/replication_record"}},"created_at":"2026-07-05T10:44:27.722083+00:00","updated_at":"2026-07-05T10:44:27.722083+00:00"}