{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2024:B2CTDRRNCN6Y43ADCKI7WUIP34","short_pith_number":"pith:B2CTDRRN","schema_version":"1.0","canonical_sha256":"0e8531c62d137d8e6c031291fb510fdf02576b550cd096d314e60df1d94bf899","source":{"kind":"arxiv","id":"2403.10553","version":1},"attestation_state":"computed","paper":{"title":"Learning to Watermark LLM-generated Text via Reinforcement Learning","license":"http://creativecommons.org/licenses/by/4.0/","headline":"","cross_cats":["cs.AI","cs.CR"],"primary_cat":"cs.LG","authors_text":"Xiaojun Xu, Yang Liu, Yuanshun Yao","submitted_at":"2024-03-13T03:43:39Z","abstract_excerpt":"We study how to watermark LLM outputs, i.e. embedding algorithmically detectable signals into LLM-generated text to track misuse. Unlike the current mainstream methods that work with a fixed LLM, we expand the watermark design space by including the LLM tuning stage in the watermark pipeline. While prior works focus on token-level watermark that embeds signals into the output, we design a model-level watermark that embeds signals into the LLM weights, and such signals can be detected by a paired detector. We propose a co-training framework based on reinforcement learning that iteratively (1) t"},"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":"2403.10553","kind":"arxiv","version":1},"metadata":{"license":"http://creativecommons.org/licenses/by/4.0/","primary_cat":"cs.LG","submitted_at":"2024-03-13T03:43:39Z","cross_cats_sorted":["cs.AI","cs.CR"],"title_canon_sha256":"fe840966481b77b4e40c6a58d23562be62ce49a2c51fe700e3bc90104854c85c","abstract_canon_sha256":"baf30a0275d1ba23b23199abc349f96ab49444f162e506dbb406e5f27449cad9"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-07-05T07:57:02.969717Z","signature_b64":"OEjhH2alk+MVxX9tHfIzNf/8wzMgCDGukbW3Tam7Zng2hqihqG1d0aj/PWAUw9WO9KMWAfZhF46+Qqk9gqbYBw==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"0e8531c62d137d8e6c031291fb510fdf02576b550cd096d314e60df1d94bf899","last_reissued_at":"2026-07-05T07:57:02.969308Z","signature_status":"signed_v1","first_computed_at":"2026-07-05T07:57:02.969308Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"Learning to Watermark LLM-generated Text via Reinforcement Learning","license":"http://creativecommons.org/licenses/by/4.0/","headline":"","cross_cats":["cs.AI","cs.CR"],"primary_cat":"cs.LG","authors_text":"Xiaojun Xu, Yang Liu, Yuanshun Yao","submitted_at":"2024-03-13T03:43:39Z","abstract_excerpt":"We study how to watermark LLM outputs, i.e. embedding algorithmically detectable signals into LLM-generated text to track misuse. Unlike the current mainstream methods that work with a fixed LLM, we expand the watermark design space by including the LLM tuning stage in the watermark pipeline. While prior works focus on token-level watermark that embeds signals into the output, we design a model-level watermark that embeds signals into the LLM weights, and such signals can be detected by a paired detector. We propose a co-training framework based on reinforcement learning that iteratively (1) t"},"claims":{"count":0,"items":[],"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"source":{"id":"2403.10553","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/2403.10553/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":"2403.10553","created_at":"2026-07-05T07:57:02.969383+00:00"},{"alias_kind":"arxiv_version","alias_value":"2403.10553v1","created_at":"2026-07-05T07:57:02.969383+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2403.10553","created_at":"2026-07-05T07:57:02.969383+00:00"},{"alias_kind":"pith_short_12","alias_value":"B2CTDRRNCN6Y","created_at":"2026-07-05T07:57:02.969383+00:00"},{"alias_kind":"pith_short_16","alias_value":"B2CTDRRNCN6Y43AD","created_at":"2026-07-05T07:57:02.969383+00:00"},{"alias_kind":"pith_short_8","alias_value":"B2CTDRRN","created_at":"2026-07-05T07:57:02.969383+00:00"}],"events":[],"event_summary":{},"paper_claims":[],"inbound_citations":{"count":5,"internal_anchor_count":0,"sample":[{"citing_arxiv_id":"2508.11548","citing_title":"Copyright Protection for Large Language Models: A Survey of Methods, Challenges, and Trends","ref_index":159,"is_internal_anchor":false},{"citing_arxiv_id":"2510.18333","citing_title":"Position: LLM Watermarking Should Align Stakeholders' Incentives for Practical Adoption","ref_index":60,"is_internal_anchor":false},{"citing_arxiv_id":"2605.08826","citing_title":"Fundamental Trade-Offs in Multi-Bit Watermarking of Stochastic Processes","ref_index":31,"is_internal_anchor":false},{"citing_arxiv_id":"2604.10893","citing_title":"Beyond A Fixed Seal: Adaptive Stealing Watermark in Large Language Models","ref_index":8,"is_internal_anchor":false},{"citing_arxiv_id":"2604.08759","citing_title":"Optimal Multi-bit Generative Watermarking Schemes Under Worst-Case False-Alarm Constraints","ref_index":3,"is_internal_anchor":false}]},"formal_canon":{"evidence_count":0,"sample":[],"anchors":[]},"links":{"html":"https://pith.science/pith/B2CTDRRNCN6Y43ADCKI7WUIP34","json":"https://pith.science/pith/B2CTDRRNCN6Y43ADCKI7WUIP34.json","graph_json":"https://pith.science/api/pith-number/B2CTDRRNCN6Y43ADCKI7WUIP34/graph.json","events_json":"https://pith.science/api/pith-number/B2CTDRRNCN6Y43ADCKI7WUIP34/events.json","paper":"https://pith.science/paper/B2CTDRRN"},"agent_actions":{"view_html":"https://pith.science/pith/B2CTDRRNCN6Y43ADCKI7WUIP34","download_json":"https://pith.science/pith/B2CTDRRNCN6Y43ADCKI7WUIP34.json","view_paper":"https://pith.science/paper/B2CTDRRN","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2403.10553&json=true","fetch_graph":"https://pith.science/api/pith-number/B2CTDRRNCN6Y43ADCKI7WUIP34/graph.json","fetch_events":"https://pith.science/api/pith-number/B2CTDRRNCN6Y43ADCKI7WUIP34/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/B2CTDRRNCN6Y43ADCKI7WUIP34/action/timestamp_anchor","attest_storage":"https://pith.science/pith/B2CTDRRNCN6Y43ADCKI7WUIP34/action/storage_attestation","attest_author":"https://pith.science/pith/B2CTDRRNCN6Y43ADCKI7WUIP34/action/author_attestation","sign_citation":"https://pith.science/pith/B2CTDRRNCN6Y43ADCKI7WUIP34/action/citation_signature","submit_replication":"https://pith.science/pith/B2CTDRRNCN6Y43ADCKI7WUIP34/action/replication_record"}},"created_at":"2026-07-05T07:57:02.969383+00:00","updated_at":"2026-07-05T07:57:02.969383+00:00"}