{"bundle_type":"pith_open_graph_bundle","bundle_version":"1.0","pith_number":"pith:2026:CLP3G3KJCWLHLAL2GB7NDVZ6L5","short_pith_number":"pith:CLP3G3KJ","canonical_record":{"source":{"id":"2601.03190","kind":"arxiv","version":4},"metadata":{"license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","primary_cat":"cs.CL","submitted_at":"2026-01-06T17:10:48Z","cross_cats_sorted":[],"title_canon_sha256":"e6678e9e470c2e43f715bb355b245391abc241348cacfd4e1d4450277c717d6b","abstract_canon_sha256":"c2ed23206ed1bb7645ed04853644862aafd8e9912dcdf922f6665b9f7d82cbea"},"schema_version":"1.0"},"canonical_sha256":"12dfb36d49159675817a307ed1d73e5f446e95fe9f66d59222127dec4e569f37","source":{"kind":"arxiv","id":"2601.03190","version":4},"source_aliases":[{"alias_kind":"arxiv","alias_value":"2601.03190","created_at":"2026-07-27T00:19:45Z"},{"alias_kind":"arxiv_version","alias_value":"2601.03190v4","created_at":"2026-07-27T00:19:45Z"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2601.03190","created_at":"2026-07-27T00:19:45Z"},{"alias_kind":"pith_short_12","alias_value":"CLP3G3KJCWLH","created_at":"2026-07-27T00:19:45Z"},{"alias_kind":"pith_short_16","alias_value":"CLP3G3KJCWLHLAL2","created_at":"2026-07-27T00:19:45Z"},{"alias_kind":"pith_short_8","alias_value":"CLP3G3KJ","created_at":"2026-07-27T00:19:45Z"}],"events":[{"event_type":"record_created","subject_pith_number":"pith:2026:CLP3G3KJCWLHLAL2GB7NDVZ6L5","target":"record","payload":{"canonical_record":{"source":{"id":"2601.03190","kind":"arxiv","version":4},"metadata":{"license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","primary_cat":"cs.CL","submitted_at":"2026-01-06T17:10:48Z","cross_cats_sorted":[],"title_canon_sha256":"e6678e9e470c2e43f715bb355b245391abc241348cacfd4e1d4450277c717d6b","abstract_canon_sha256":"c2ed23206ed1bb7645ed04853644862aafd8e9912dcdf922f6665b9f7d82cbea"},"schema_version":"1.0"},"canonical_sha256":"12dfb36d49159675817a307ed1d73e5f446e95fe9f66d59222127dec4e569f37","receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-07-27T00:19:45.913798Z","signature_b64":"6nxKvH2AbzQc97uIyiVc+qV6Byz8td1WGkF8ncnbIvflzoAoVjT0oEeJYAJSB16a2PY3wr+DNm0brk+snPZODQ==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"12dfb36d49159675817a307ed1d73e5f446e95fe9f66d59222127dec4e569f37","last_reissued_at":"2026-07-27T00:19:45.912777Z","signature_status":"signed_v1","first_computed_at":"2026-07-27T00:19:45.912777Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"source_kind":"arxiv","source_id":"2601.03190","source_version":4,"attestation_state":"computed"},"signer":{"signer_id":"pith.science","signer_type":"pith_registry","key_id":"pith-v1-2026-05","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"created_at":"2026-07-27T00:19:45Z","supersedes":[],"prev_event":null,"signature":{"signature_status":"signed_v1","algorithm":"ed25519","key_id":"pith-v1-2026-05","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54","signature_b64":"xubEtwkSr0BCrD0HPWKDoOh1rRMxjt9rLRrOYuBEZebE1ZxzKSef7BoCtpuWHkyIcWanIkGmTdXO84Pq/jPrAA==","signed_message":"open_graph_event_sha256_bytes","signed_at":"2026-08-04T23:07:38.272481Z"},"content_sha256":"4d12033f2a31642129fbc7c92562cf04a67404044f176881a0dfb3c22ecb86be","schema_version":"1.0","event_id":"sha256:4d12033f2a31642129fbc7c92562cf04a67404044f176881a0dfb3c22ecb86be"},{"event_type":"graph_snapshot","subject_pith_number":"pith:2026:CLP3G3KJCWLHLAL2GB7NDVZ6L5","target":"graph","payload":{"graph_snapshot":{"paper":{"title":"Maximizing Local Entropy Where It Matters: Prefix-Aware Localized LLM Unlearning","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","headline":"Suppressing only the sensitive prefix and flattening top-k logits suffices to unlearn specific sequences in LLMs.","cross_cats":[],"primary_cat":"cs.CL","authors_text":"Binbin Zheng, Fei Shen, Long Bai, Naixin Zhai, Pengyang Shao, Xun Yang, Yonghui Yang","submitted_at":"2026-01-06T17:10:48Z","abstract_excerpt":"Machine unlearning aims to forget sensitive knowledge from Large Language Models (LLMs) while maintaining general utility. However, existing approaches typically treat all tokens in a response indiscriminately and enforce uncertainty over the entire vocabulary. This global treatment results in unnecessary utility degradation and extends optimization to content-agnostic regions. To address these limitations, we propose PALU (Prefix-Aware Localized Unlearning), a framework driven by a local entropy maximization objective across both temporal and vocabulary dimensions. PALU reveals that (i) suppr"},"claims":{"count":4,"items":[{"kind":"strongest_claim","text":"suppressing the sensitive prefix alone is sufficient to sever the causal generation link, and flattening only the top-k logits is adequate to maximize uncertainty in the critical subspace","source":"verdict.strongest_claim","status":"machine_extracted","claim_id":"C1","attestation":"unclaimed"},{"kind":"weakest_assumption","text":"That restricting entropy maximization to the prefix and top-k subspace will reliably break the generation of the full sensitive sequence without side effects on unrelated content, as claimed from the two findings (i) and (ii).","source":"verdict.weakest_assumption","status":"machine_extracted","claim_id":"C2","attestation":"unclaimed"},{"kind":"one_line_summary","text":"PALU improves LLM unlearning by restricting entropy maximization to sensitive prefixes and top-k logits, achieving better forgetting with less utility loss.","source":"verdict.one_line_summary","status":"machine_extracted","claim_id":"C3","attestation":"unclaimed"},{"kind":"headline","text":"Suppressing only the sensitive prefix and flattening top-k logits suffices to unlearn specific sequences in LLMs.","source":"verdict.pith_extraction.headline","status":"machine_extracted","claim_id":"C4","attestation":"unclaimed"}],"snapshot_sha256":"299a1d07d177cb9699f45e5e202235dd29db22ddb0f7440c17a2e6725252347d"},"source":{"id":"2601.03190","kind":"arxiv","version":4},"verdict":{"id":"064467f2-9a0e-48ed-b910-af153bb02f5f","model_set":{"reader":"grok-4.3"},"created_at":"2026-05-16T16:43:22.470762Z","strongest_claim":"suppressing the sensitive prefix alone is sufficient to sever the causal generation link, and flattening only the top-k logits is adequate to maximize uncertainty in the critical subspace","one_line_summary":"PALU improves LLM unlearning by restricting entropy maximization to sensitive prefixes and top-k logits, achieving better forgetting with less utility loss.","pipeline_version":"pith-pipeline@v0.9.0","weakest_assumption":"That restricting entropy maximization to the prefix and top-k subspace will reliably break the generation of the full sensitive sequence without side effects on unrelated content, as claimed from the two findings (i) and (ii).","pith_extraction_headline":"Suppressing only the sensitive prefix and flattening top-k logits suffices to unlearn specific sequences in LLMs."},"integrity":{"clean":true,"summary":{"advisory":0,"critical":0,"by_detector":{},"informational":0},"endpoint":"/pith/2601.03190/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":2,"snapshot_sha256":"505344e5d004d07d50fbc0c1dffda1ac7fcf413449b9b85b086e1b3c1a4e4812"},"author_claims":{"count":0,"strong_count":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"builder_version":"pith-number-builder-2026-05-17-v1"},"verdict_id":"064467f2-9a0e-48ed-b910-af153bb02f5f"},"signer":{"signer_id":"pith.science","signer_type":"pith_registry","key_id":"pith-v1-2026-05","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"created_at":"2026-07-27T00:19:45Z","supersedes":[],"prev_event":null,"signature":{"signature_status":"signed_v1","algorithm":"ed25519","key_id":"pith-v1-2026-05","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54","signature_b64":"jNTfqKPRAUNh32bwTWvpmlMSZCE8aQHk9n5i/mo3jcU1836xvQ1TG05KK7boaE6tJM+0oQn4oDUGwUmS7lZoAQ==","signed_message":"open_graph_event_sha256_bytes","signed_at":"2026-08-04T23:07:38.278940Z"},"content_sha256":"e3c701caa56a47eb5965753fd6b63346375b864d33c9117fdbe4d13179fc4c8f","schema_version":"1.0","event_id":"sha256:e3c701caa56a47eb5965753fd6b63346375b864d33c9117fdbe4d13179fc4c8f"}],"timestamp_proofs":[],"mirror_hints":[{"mirror_type":"https","name":"Pith Resolver","base_url":"https://pith.science","bundle_url":"https://pith.science/pith/CLP3G3KJCWLHLAL2GB7NDVZ6L5/bundle.json","state_url":"https://pith.science/pith/CLP3G3KJCWLHLAL2GB7NDVZ6L5/state.json","well_known_bundle_url":"https://pith.science/.well-known/pith/CLP3G3KJCWLHLAL2GB7NDVZ6L5/bundle.json","status":"primary"}],"public_keys":[{"key_id":"pith-v1-2026-05","algorithm":"ed25519","format":"raw","public_key_b64":"stVStoiQhXFxp4s2pdzPNoqVNBMojDU/fJ2db5S3CbM=","public_key_hex":"b2d552b68890857171a78b36a5dccf368a953413288c353f7c9d9d6f94b709b3","fingerprint_sha256_b32_first128bits":"RVFV5Z2OI2J3ZUO7ERDEBCYNKS","fingerprint_sha256_hex":"8d4b5ee74e4693bcd1df2446408b0d54","rotates_at":null,"url":"https://pith.science/pith-signing-key.json","notes":"Pith uses this Ed25519 key to sign canonical record SHA-256 digests. Verify with: ed25519_verify(public_key, message=canonical_sha256_bytes, signature=base64decode(signature_b64))."}],"merge_version":"pith-open-graph-merge-v1","built_at":"2026-08-04T23:07:38Z","links":{"resolver":"https://pith.science/pith/CLP3G3KJCWLHLAL2GB7NDVZ6L5","bundle":"https://pith.science/pith/CLP3G3KJCWLHLAL2GB7NDVZ6L5/bundle.json","state":"https://pith.science/pith/CLP3G3KJCWLHLAL2GB7NDVZ6L5/state.json","well_known_bundle":"https://pith.science/.well-known/pith/CLP3G3KJCWLHLAL2GB7NDVZ6L5/bundle.json"},"state":{"state_type":"pith_open_graph_state","state_version":"1.0","pith_number":"pith:2026:CLP3G3KJCWLHLAL2GB7NDVZ6L5","merge_version":"pith-open-graph-merge-v1","event_count":2,"valid_event_count":2,"invalid_event_count":0,"equivocation_count":0,"current":{"canonical_record":{"metadata":{"abstract_canon_sha256":"c2ed23206ed1bb7645ed04853644862aafd8e9912dcdf922f6665b9f7d82cbea","cross_cats_sorted":[],"license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","primary_cat":"cs.CL","submitted_at":"2026-01-06T17:10:48Z","title_canon_sha256":"e6678e9e470c2e43f715bb355b245391abc241348cacfd4e1d4450277c717d6b"},"schema_version":"1.0","source":{"id":"2601.03190","kind":"arxiv","version":4}},"source_aliases":[{"alias_kind":"arxiv","alias_value":"2601.03190","created_at":"2026-07-27T00:19:45Z"},{"alias_kind":"arxiv_version","alias_value":"2601.03190v4","created_at":"2026-07-27T00:19:45Z"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2601.03190","created_at":"2026-07-27T00:19:45Z"},{"alias_kind":"pith_short_12","alias_value":"CLP3G3KJCWLH","created_at":"2026-07-27T00:19:45Z"},{"alias_kind":"pith_short_16","alias_value":"CLP3G3KJCWLHLAL2","created_at":"2026-07-27T00:19:45Z"},{"alias_kind":"pith_short_8","alias_value":"CLP3G3KJ","created_at":"2026-07-27T00:19:45Z"}],"graph_snapshots":[{"event_id":"sha256:e3c701caa56a47eb5965753fd6b63346375b864d33c9117fdbe4d13179fc4c8f","target":"graph","created_at":"2026-07-27T00:19:45Z","signer":{"key_id":"pith-v1-2026-05","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54","signer_id":"pith.science","signer_type":"pith_registry"},"payload":{"graph_snapshot":{"author_claims":{"count":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57","strong_count":0},"builder_version":"pith-number-builder-2026-05-17-v1","claims":{"count":4,"items":[{"attestation":"unclaimed","claim_id":"C1","kind":"strongest_claim","source":"verdict.strongest_claim","status":"machine_extracted","text":"suppressing the sensitive prefix alone is sufficient to sever the causal generation link, and flattening only the top-k logits is adequate to maximize uncertainty in the critical subspace"},{"attestation":"unclaimed","claim_id":"C2","kind":"weakest_assumption","source":"verdict.weakest_assumption","status":"machine_extracted","text":"That restricting entropy maximization to the prefix and top-k subspace will reliably break the generation of the full sensitive sequence without side effects on unrelated content, as claimed from the two findings (i) and (ii)."},{"attestation":"unclaimed","claim_id":"C3","kind":"one_line_summary","source":"verdict.one_line_summary","status":"machine_extracted","text":"PALU improves LLM unlearning by restricting entropy maximization to sensitive prefixes and top-k logits, achieving better forgetting with less utility loss."},{"attestation":"unclaimed","claim_id":"C4","kind":"headline","source":"verdict.pith_extraction.headline","status":"machine_extracted","text":"Suppressing only the sensitive prefix and flattening top-k logits suffices to unlearn specific sequences in LLMs."}],"snapshot_sha256":"299a1d07d177cb9699f45e5e202235dd29db22ddb0f7440c17a2e6725252347d"},"formal_canon":{"evidence_count":2,"snapshot_sha256":"505344e5d004d07d50fbc0c1dffda1ac7fcf413449b9b85b086e1b3c1a4e4812"},"integrity":{"available":true,"clean":true,"detectors_run":[],"endpoint":"/pith/2601.03190/integrity.json","findings":[],"snapshot_sha256":"c28c3603d3b5d939e8dc4c7e95fa8dfce3d595e45f758748cecf8e644a296938","summary":{"advisory":0,"by_detector":{},"critical":0,"informational":0}},"paper":{"abstract_excerpt":"Machine unlearning aims to forget sensitive knowledge from Large Language Models (LLMs) while maintaining general utility. However, existing approaches typically treat all tokens in a response indiscriminately and enforce uncertainty over the entire vocabulary. This global treatment results in unnecessary utility degradation and extends optimization to content-agnostic regions. To address these limitations, we propose PALU (Prefix-Aware Localized Unlearning), a framework driven by a local entropy maximization objective across both temporal and vocabulary dimensions. PALU reveals that (i) suppr","authors_text":"Binbin Zheng, Fei Shen, Long Bai, Naixin Zhai, Pengyang Shao, Xun Yang, Yonghui Yang","cross_cats":[],"headline":"Suppressing only the sensitive prefix and flattening top-k logits suffices to unlearn specific sequences in LLMs.","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","primary_cat":"cs.CL","submitted_at":"2026-01-06T17:10:48Z","title":"Maximizing Local Entropy Where It Matters: Prefix-Aware Localized LLM Unlearning"},"references":{"count":0,"internal_anchors":0,"resolved_work":0,"sample":[],"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"source":{"id":"2601.03190","kind":"arxiv","version":4},"verdict":{"created_at":"2026-05-16T16:43:22.470762Z","id":"064467f2-9a0e-48ed-b910-af153bb02f5f","model_set":{"reader":"grok-4.3"},"one_line_summary":"PALU improves LLM unlearning by restricting entropy maximization to sensitive prefixes and top-k logits, achieving better forgetting with less utility loss.","pipeline_version":"pith-pipeline@v0.9.0","pith_extraction_headline":"Suppressing only the sensitive prefix and flattening top-k logits suffices to unlearn specific sequences in LLMs.","strongest_claim":"suppressing the sensitive prefix alone is sufficient to sever the causal generation link, and flattening only the top-k logits is adequate to maximize uncertainty in the critical subspace","weakest_assumption":"That restricting entropy maximization to the prefix and top-k subspace will reliably break the generation of the full sensitive sequence without side effects on unrelated content, as claimed from the two findings (i) and (ii)."}},"verdict_id":"064467f2-9a0e-48ed-b910-af153bb02f5f"}}],"author_attestations":[],"timestamp_anchors":[],"storage_attestations":[],"citation_signatures":[],"replication_records":[],"corrections":[],"mirror_hints":[],"record_created":{"event_id":"sha256:4d12033f2a31642129fbc7c92562cf04a67404044f176881a0dfb3c22ecb86be","target":"record","created_at":"2026-07-27T00:19:45Z","signer":{"key_id":"pith-v1-2026-05","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54","signer_id":"pith.science","signer_type":"pith_registry"},"payload":{"attestation_state":"computed","canonical_record":{"metadata":{"abstract_canon_sha256":"c2ed23206ed1bb7645ed04853644862aafd8e9912dcdf922f6665b9f7d82cbea","cross_cats_sorted":[],"license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","primary_cat":"cs.CL","submitted_at":"2026-01-06T17:10:48Z","title_canon_sha256":"e6678e9e470c2e43f715bb355b245391abc241348cacfd4e1d4450277c717d6b"},"schema_version":"1.0","source":{"id":"2601.03190","kind":"arxiv","version":4}},"canonical_sha256":"12dfb36d49159675817a307ed1d73e5f446e95fe9f66d59222127dec4e569f37","receipt":{"algorithm":"ed25519","builder_version":"pith-number-builder-2026-05-17-v1","canonical_sha256":"12dfb36d49159675817a307ed1d73e5f446e95fe9f66d59222127dec4e569f37","first_computed_at":"2026-07-27T00:19:45.912777Z","key_id":"pith-v1-2026-05","kind":"pith_receipt","last_reissued_at":"2026-07-27T00:19:45.912777Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54","receipt_version":"0.3","signature_b64":"6nxKvH2AbzQc97uIyiVc+qV6Byz8td1WGkF8ncnbIvflzoAoVjT0oEeJYAJSB16a2PY3wr+DNm0brk+snPZODQ==","signature_status":"signed_v1","signed_at":"2026-07-27T00:19:45.913798Z","signed_message":"canonical_sha256_bytes"},"source_id":"2601.03190","source_kind":"arxiv","source_version":4}}},"equivocations":[],"invalid_events":[],"applied_event_ids":["sha256:4d12033f2a31642129fbc7c92562cf04a67404044f176881a0dfb3c22ecb86be","sha256:e3c701caa56a47eb5965753fd6b63346375b864d33c9117fdbe4d13179fc4c8f"],"state_sha256":"ecf6f60b5fe62e06ffa1d437a04d32d14b4dfd4d90e57e2225dd604fe126e2da"},"bundle_signature":{"signature_status":"signed_v1","algorithm":"ed25519","key_id":"pith-v1-2026-05","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54","signature_b64":"WZaUahjasX6fLrVT0/n6/v4GZ4rK6gVH2ijTs67LNu6xwj1gBU0BGbslXmgSnqFQ1nLZorCs8FdDBiT+esuvBA==","signed_message":"bundle_sha256_bytes","signed_at":"2026-08-04T23:07:38.286417Z","bundle_sha256":"397af48fd17e82506b98186aefcc8aef57d5dbd144666ac0c76884df300b3072"}}