{"paper":{"title":"LiSA: Lifelong Safety Adaptation via Conservative Policy Induction","license":"http://creativecommons.org/licenses/by/4.0/","headline":"LiSA lets fixed guardrails adapt to sparse noisy user feedback by inducing conservative reusable policies.","cross_cats":["cs.CL","cs.CR"],"primary_cat":"cs.LG","authors_text":"Bharath Chandrasekhar, Bhavana Dalvi Mishra, Kyomin Jung, Lesly Miculicich, Long T. Le, Mihir Parmar, Minbeom Kim, Phillip Wallis, Tomas Pfister","submitted_at":"2026-05-14T06:47:35Z","abstract_excerpt":"As AI agents move from chat interfaces to systems that read private data, call tools, and execute multi-step workflows, guardrails become a last line of defense against concrete deployment harms. In these settings, guardrail failures are no longer merely answer-quality errors: they can leak secrets, authorize unsafe actions, or block legitimate work. The hardest failures are often contextual: whether an action is acceptable depends on local privacy norms, organizational policies, and user expectations that resist pre-deployment specification. This creates a practical gap: guardrails must adapt"},"claims":{"count":4,"items":[{"kind":"strongest_claim","text":"Across PrivacyLens+, ConFaide+, and AgentHarm, LiSA consistently outperforms strong memory-based baselines under sparse feedback, remains robust under noisy user feedback even at 20% label-flip rates, and pushes the latency--performance frontier beyond backbone model scaling.","source":"verdict.strongest_claim","status":"machine_extracted","claim_id":"C1","attestation":"unclaimed"},{"kind":"weakest_assumption","text":"That occasional sparse and noisy user-reported failures can be reliably converted into reusable policy abstractions that generalize without overgeneralization, supported by conflict-aware local rules and evidence-aware posterior lower-bound gating.","source":"verdict.weakest_assumption","status":"machine_extracted","claim_id":"C2","attestation":"unclaimed"},{"kind":"one_line_summary","text":"LiSA improves AI guardrails lifelong by inducing conservative policies from sparse noisy failure reports via structured memory, conflict-aware rules, and posterior lower-bound gating.","source":"verdict.one_line_summary","status":"machine_extracted","claim_id":"C3","attestation":"unclaimed"},{"kind":"headline","text":"LiSA lets fixed guardrails adapt to sparse noisy user feedback by inducing conservative reusable policies.","source":"verdict.pith_extraction.headline","status":"machine_extracted","claim_id":"C4","attestation":"unclaimed"}],"snapshot_sha256":"55a87722fcbff0044106119fb620af0794a698b2a7081782e87b46d1dac290f4"},"source":{"id":"2605.14454","kind":"arxiv","version":1},"verdict":{"id":"32776a9c-4fa9-47e8-b3c0-076c96f06707","model_set":{"reader":"grok-4.3"},"created_at":"2026-05-15T01:49:58.412235Z","strongest_claim":"Across PrivacyLens+, ConFaide+, and AgentHarm, LiSA consistently outperforms strong memory-based baselines under sparse feedback, remains robust under noisy user feedback even at 20% label-flip rates, and pushes the latency--performance frontier beyond backbone model scaling.","one_line_summary":"LiSA improves AI guardrails lifelong by inducing conservative policies from sparse noisy failure reports via structured memory, conflict-aware rules, and posterior lower-bound gating.","pipeline_version":"pith-pipeline@v0.9.0","weakest_assumption":"That occasional sparse and noisy user-reported failures can be reliably converted into reusable policy abstractions that generalize without overgeneralization, supported by conflict-aware local rules and evidence-aware posterior lower-bound gating.","pith_extraction_headline":"LiSA lets fixed guardrails adapt to sparse noisy user feedback by inducing conservative reusable policies."},"references":{"count":47,"sample":[{"doi":"","year":2004,"title":"Privacy as contextual integrity , author=. Wash. L. Rev. , volume=. 2004 , publisher=","work_id":"8b6785e7-0848-443c-9c81-50e27067b8a6","ref_index":1,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":null,"title":"Gemini Embedding: Generalizable Embeddings from Gemini","work_id":"911b6918-a128-453f-ae99-94388c38fcb1","ref_index":2,"cited_arxiv_id":"2503.07891","is_internal_anchor":true},{"doi":"","year":null,"title":"Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , pages=","work_id":"3dc6599a-4158-4b45-8b6c-3946d23d4ba9","ref_index":3,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2025,"title":"Introducing Claude Haiku 4.5 , author=. 2025 , url=","work_id":"c42de211-a529-4fe2-8aeb-bdfede431e1e","ref_index":4,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":null,"title":"A new era of intelligence with Gemini 3 , author=. Google. URL: https://blog.google/products-and-platforms/products/gemini/gemini-3 , year=","work_id":"7e33519b-a76e-4d28-8f8a-c86ae2704c25","ref_index":5,"cited_arxiv_id":"","is_internal_anchor":false}],"resolved_work":47,"snapshot_sha256":"b41929fcfca4c753cb2366db7b215e4b27167ecca3a63af89546f594c96c8d74","internal_anchors":5},"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"}