{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2026:3BT6KYVMF5JWIDWL3YPTU4PN5Q","short_pith_number":"pith:3BT6KYVM","schema_version":"1.0","canonical_sha256":"d867e562ac2f53640ecbde1f3a71edec3e23cfb031c3fe0d3256efdc1d6db991","source":{"kind":"arxiv","id":"2604.21251","version":5},"attestation_state":"computed","paper":{"title":"CAP: Controllable Alignment Prompting for Unlearning in LLMs","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","headline":"Reinforcement learning trains prompts that suppress specific knowledge in fixed LLMs while preserving general capabilities and allowing reversal by prompt removal.","cross_cats":["cs.AI"],"primary_cat":"cs.LG","authors_text":"Guangchun Luo, Hongli Pu, Jie Ou, Jingwen Pu, Jinyu Guo, Meng Yang, Wenhong Tian, Wenyi Li, Xunlei Chen, Zhaokun Wang","submitted_at":"2026-04-23T03:42:41Z","abstract_excerpt":"Large language models (LLMs) trained on unfiltered corpora inherently risk retaining sensitive information, necessitating selective knowledge unlearning for regulatory compliance and ethical safety. However, existing parameter-modifying methods face fundamental limitations: high computational costs, uncontrollable forgetting boundaries, and strict dependency on model weight access. These constraints render them impractical for closed-source models, yet current non-invasive alternatives remain unsystematic and reliant on empirical experience. To address these challenges, we propose the Controll"},"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":true,"formal_links_present":true},"canonical_record":{"source":{"id":"2604.21251","kind":"arxiv","version":5},"metadata":{"license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","primary_cat":"cs.LG","submitted_at":"2026-04-23T03:42:41Z","cross_cats_sorted":["cs.AI"],"title_canon_sha256":"dc58c401d2e10dea15f01850d00bfb627f85b25b72306526e973e7d893b49b4a","abstract_canon_sha256":"4b30d881f1cf3332097f0dc777d05f6157b3430ab07a46e1d77f206d6658f781"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-05-20T00:00:39.304387Z","signature_b64":"PjaJ8OPyeOAy8i0omKL+4DWos2whrjFLaDORD/MBskztKJqgjECdrLDrrf/TAx6d6SJxEd2dou2Mv5SwX3eDBw==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"d867e562ac2f53640ecbde1f3a71edec3e23cfb031c3fe0d3256efdc1d6db991","last_reissued_at":"2026-05-20T00:00:39.303678Z","signature_status":"signed_v1","first_computed_at":"2026-05-20T00:00:39.303678Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"CAP: Controllable Alignment Prompting for Unlearning in LLMs","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","headline":"Reinforcement learning trains prompts that suppress specific knowledge in fixed LLMs while preserving general capabilities and allowing reversal by prompt removal.","cross_cats":["cs.AI"],"primary_cat":"cs.LG","authors_text":"Guangchun Luo, Hongli Pu, Jie Ou, Jingwen Pu, Jinyu Guo, Meng Yang, Wenhong Tian, Wenyi Li, Xunlei Chen, Zhaokun Wang","submitted_at":"2026-04-23T03:42:41Z","abstract_excerpt":"Large language models (LLMs) trained on unfiltered corpora inherently risk retaining sensitive information, necessitating selective knowledge unlearning for regulatory compliance and ethical safety. However, existing parameter-modifying methods face fundamental limitations: high computational costs, uncontrollable forgetting boundaries, and strict dependency on model weight access. These constraints render them impractical for closed-source models, yet current non-invasive alternatives remain unsystematic and reliant on empirical experience. To address these challenges, we propose the Controll"},"claims":{"count":4,"items":[{"kind":"strongest_claim","text":"CAP achieves precise, controllable unlearning without updating model parameters, establishing a dynamic alignment mechanism that overcomes the transferability limitations of prior methods.","source":"verdict.strongest_claim","status":"machine_extracted","claim_id":"C1","attestation":"unclaimed"},{"kind":"weakest_assumption","text":"That reinforcement learning can train a prompt generator to collaborate with a fixed LLM such that target knowledge is suppressed while general capabilities remain selectively preserved and the effect is reversible upon prompt revocation.","source":"verdict.weakest_assumption","status":"machine_extracted","claim_id":"C2","attestation":"unclaimed"},{"kind":"one_line_summary","text":"CAP is a reinforcement-learning-driven prompt optimization framework that suppresses target knowledge in LLMs while preserving general capabilities, enabling reversible unlearning without any parameter updates.","source":"verdict.one_line_summary","status":"machine_extracted","claim_id":"C3","attestation":"unclaimed"},{"kind":"headline","text":"Reinforcement learning trains prompts that suppress specific knowledge in fixed LLMs while preserving general capabilities and allowing reversal by prompt removal.","source":"verdict.pith_extraction.headline","status":"machine_extracted","claim_id":"C4","attestation":"unclaimed"}],"snapshot_sha256":"8c389cb8520bffc1dc4f487684b4f64f2cae6155eb8d3a2e3f347e737b49f2f0"},"source":{"id":"2604.21251","kind":"arxiv","version":5},"verdict":{"id":"122d6b06-2259-48eb-8386-3ea584f94e1a","model_set":{"reader":"grok-4.3"},"created_at":"2026-05-19T17:09:46.710788Z","strongest_claim":"CAP achieves precise, controllable unlearning without updating model parameters, establishing a dynamic alignment mechanism that overcomes the transferability limitations of prior methods.","one_line_summary":"CAP is a reinforcement-learning-driven prompt optimization framework that suppresses target knowledge in LLMs while preserving general capabilities, enabling reversible unlearning without any parameter updates.","pipeline_version":"pith-pipeline@v0.9.0","weakest_assumption":"That reinforcement learning can train a prompt generator to collaborate with a fixed LLM such that target knowledge is suppressed while general capabilities remain selectively preserved and the effect is reversible upon prompt revocation.","pith_extraction_headline":"Reinforcement learning trains prompts that suppress specific knowledge in fixed LLMs while preserving general capabilities and allowing reversal by prompt removal."},"integrity":{"clean":true,"summary":{"advisory":0,"critical":0,"by_detector":{},"informational":0},"endpoint":"/pith/2604.21251/integrity.json","findings":[],"available":true,"detectors_run":[],"snapshot_sha256":"c28c3603d3b5d939e8dc4c7e95fa8dfce3d595e45f758748cecf8e644a296938"},"references":{"count":9,"sample":[{"doi":"","year":2024,"title":"Yi: Open Foundation Models by 01.AI","work_id":"8efee8a1-5e3c-4851-9c65-18e3d1d9e769","ref_index":1,"cited_arxiv_id":"2403.04652","is_internal_anchor":true},{"doi":"","year":2024,"title":"DeepSeek-V3 Technical Report","work_id":"57d2791d-2219-4c31-a077-afc04b12a75c","ref_index":2,"cited_arxiv_id":"2412.19437","is_internal_anchor":true},{"doi":"","year":2024,"title":"ChatGLM: A Family of Large Language Models from GLM-130B to GLM-4 All Tools","work_id":"de9ce5af-0d8d-4b94-9793-64968d9bc06d","ref_index":3,"cited_arxiv_id":"2406.12793","is_internal_anchor":true},{"doi":"","year":2025,"title":"The WMDP Benchmark: Measuring and Reducing Malicious Use With Unlearning","work_id":"d05f8523-8089-4fdb-9c07-463952166528","ref_index":4,"cited_arxiv_id":"2403.03218","is_internal_anchor":true},{"doi":"","year":2024,"title":"TDA-RC: Task-Driven Alignment for Knowledge-Based Reasoning Chains in Large Language Models","work_id":"77051665-d34f-4db0-a66e-25fc43863478","ref_index":5,"cited_arxiv_id":"2604.04942","is_internal_anchor":true}],"resolved_work":9,"snapshot_sha256":"a7d4ae06d7fbf1248f65e8de2049f1d4782782b83cf652dae6404d9b9d53c360","internal_anchors":5},"formal_canon":{"evidence_count":2,"snapshot_sha256":"f999160140984ad363221ef326d77ff35166d22a5eb74c9407ce8cf6c209c2c9"},"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":"2604.21251","created_at":"2026-05-20T00:00:39.303791+00:00"},{"alias_kind":"arxiv_version","alias_value":"2604.21251v5","created_at":"2026-05-20T00:00:39.303791+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2604.21251","created_at":"2026-05-20T00:00:39.303791+00:00"},{"alias_kind":"pith_short_12","alias_value":"3BT6KYVMF5JW","created_at":"2026-05-20T00:00:39.303791+00:00"},{"alias_kind":"pith_short_16","alias_value":"3BT6KYVMF5JWIDWL","created_at":"2026-05-20T00:00:39.303791+00:00"},{"alias_kind":"pith_short_8","alias_value":"3BT6KYVM","created_at":"2026-05-20T00:00:39.303791+00:00"}],"events":[],"event_summary":{},"paper_claims":[],"inbound_citations":{"count":0,"internal_anchor_count":0,"sample":[]},"formal_canon":{"evidence_count":2,"sample":[],"anchors":[]},"links":{"html":"https://pith.science/pith/3BT6KYVMF5JWIDWL3YPTU4PN5Q","json":"https://pith.science/pith/3BT6KYVMF5JWIDWL3YPTU4PN5Q.json","graph_json":"https://pith.science/api/pith-number/3BT6KYVMF5JWIDWL3YPTU4PN5Q/graph.json","events_json":"https://pith.science/api/pith-number/3BT6KYVMF5JWIDWL3YPTU4PN5Q/events.json","paper":"https://pith.science/paper/3BT6KYVM"},"agent_actions":{"view_html":"https://pith.science/pith/3BT6KYVMF5JWIDWL3YPTU4PN5Q","download_json":"https://pith.science/pith/3BT6KYVMF5JWIDWL3YPTU4PN5Q.json","view_paper":"https://pith.science/paper/3BT6KYVM","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2604.21251&json=true","fetch_graph":"https://pith.science/api/pith-number/3BT6KYVMF5JWIDWL3YPTU4PN5Q/graph.json","fetch_events":"https://pith.science/api/pith-number/3BT6KYVMF5JWIDWL3YPTU4PN5Q/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/3BT6KYVMF5JWIDWL3YPTU4PN5Q/action/timestamp_anchor","attest_storage":"https://pith.science/pith/3BT6KYVMF5JWIDWL3YPTU4PN5Q/action/storage_attestation","attest_author":"https://pith.science/pith/3BT6KYVMF5JWIDWL3YPTU4PN5Q/action/author_attestation","sign_citation":"https://pith.science/pith/3BT6KYVMF5JWIDWL3YPTU4PN5Q/action/citation_signature","submit_replication":"https://pith.science/pith/3BT6KYVMF5JWIDWL3YPTU4PN5Q/action/replication_record"}},"created_at":"2026-05-20T00:00:39.303791+00:00","updated_at":"2026-05-20T00:00:39.303791+00:00"}