{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2026:L754SDXDCQQMYF35OIRBMTEVCZ","short_pith_number":"pith:L754SDXD","schema_version":"1.0","canonical_sha256":"5ffbc90ee31420cc177d7222164c95165bfb06123673e152e374c90ac87e97f7","source":{"kind":"arxiv","id":"2601.09624","version":2},"attestation_state":"computed","paper":{"title":"A Mechanistic Perspective and Circuit-Guided Difficulty Metric for Unlearning","license":"http://creativecommons.org/licenses/by/4.0/","headline":"","cross_cats":["cs.AI","cs.CL","cs.CV"],"primary_cat":"cs.LG","authors_text":"Chirag Agarwal, Hadi Amiri, Jiali Cheng, Ziheng Chen","submitted_at":"2026-01-14T16:55:58Z","abstract_excerpt":"Machine unlearning is becoming essential for building trustworthy and compliant language models. Yet unlearning success varies considerably across individual samples: some are reliably erased, while others persist despite the same procedure. We argue that this disparity is not only a data-side phenomenon, but also reflects model-internal mechanisms that encode and protect memorized information. We study this problem from a mechanistic perspective based on model circuits--structured interaction pathways that govern how predictions are formed. We propose Circuit-guided Unlearning Difficulty (CUD"},"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":"2601.09624","kind":"arxiv","version":2},"metadata":{"license":"http://creativecommons.org/licenses/by/4.0/","primary_cat":"cs.LG","submitted_at":"2026-01-14T16:55:58Z","cross_cats_sorted":["cs.AI","cs.CL","cs.CV"],"title_canon_sha256":"5f730d968844d6f6d986f9a156d9c8b28c66ecb5c4069d5cf15dcd1f960e767d","abstract_canon_sha256":"378911f82cfe19479f8b84bee59de154b17e0cd6d983beb75972605a05e155f9"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-07-28T01:23:28.349825Z","signature_b64":"w5Vb64JY1Y175iJcXHEcVOhbyKabgQ6IurBN1Skhp73b6fXnmVddPon3LxKkE0Dsh8pU8jzR6V1cwheYNquMAA==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"5ffbc90ee31420cc177d7222164c95165bfb06123673e152e374c90ac87e97f7","last_reissued_at":"2026-07-28T01:23:28.348924Z","signature_status":"signed_v1","first_computed_at":"2026-07-28T01:23:28.348924Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"A Mechanistic Perspective and Circuit-Guided Difficulty Metric for Unlearning","license":"http://creativecommons.org/licenses/by/4.0/","headline":"","cross_cats":["cs.AI","cs.CL","cs.CV"],"primary_cat":"cs.LG","authors_text":"Chirag Agarwal, Hadi Amiri, Jiali Cheng, Ziheng Chen","submitted_at":"2026-01-14T16:55:58Z","abstract_excerpt":"Machine unlearning is becoming essential for building trustworthy and compliant language models. Yet unlearning success varies considerably across individual samples: some are reliably erased, while others persist despite the same procedure. We argue that this disparity is not only a data-side phenomenon, but also reflects model-internal mechanisms that encode and protect memorized information. We study this problem from a mechanistic perspective based on model circuits--structured interaction pathways that govern how predictions are formed. We propose Circuit-guided Unlearning Difficulty (CUD"},"claims":{"count":0,"items":[],"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"source":{"id":"2601.09624","kind":"arxiv","version":2},"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/2601.09624/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":"2601.09624","created_at":"2026-07-28T01:23:28.349341+00:00"},{"alias_kind":"arxiv_version","alias_value":"2601.09624v2","created_at":"2026-07-28T01:23:28.349341+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2601.09624","created_at":"2026-07-28T01:23:28.349341+00:00"},{"alias_kind":"pith_short_12","alias_value":"L754SDXDCQQM","created_at":"2026-07-28T01:23:28.349341+00:00"},{"alias_kind":"pith_short_16","alias_value":"L754SDXDCQQMYF35","created_at":"2026-07-28T01:23:28.349341+00:00"},{"alias_kind":"pith_short_8","alias_value":"L754SDXD","created_at":"2026-07-28T01:23:28.349341+00:00"}],"events":[],"event_summary":{},"paper_claims":[],"inbound_citations":{"count":3,"internal_anchor_count":3,"sample":[{"citing_arxiv_id":"2606.07688","citing_title":"TRACER: Token ReAssignment for Concept ERasure in Generative Recommendation","ref_index":13,"is_internal_anchor":true},{"citing_arxiv_id":"2604.04982","citing_title":"CURE:Circuit-Aware Unlearning for LLM-based Recommendation","ref_index":10,"is_internal_anchor":true},{"citing_arxiv_id":"2604.18756","citing_title":"Towards Understanding the Robustness of Sparse Autoencoders","ref_index":23,"is_internal_anchor":true}]},"formal_canon":{"evidence_count":0,"sample":[],"anchors":[]},"links":{"html":"https://pith.science/pith/L754SDXDCQQMYF35OIRBMTEVCZ","json":"https://pith.science/pith/L754SDXDCQQMYF35OIRBMTEVCZ.json","graph_json":"https://pith.science/api/pith-number/L754SDXDCQQMYF35OIRBMTEVCZ/graph.json","events_json":"https://pith.science/api/pith-number/L754SDXDCQQMYF35OIRBMTEVCZ/events.json","paper":"https://pith.science/paper/L754SDXD"},"agent_actions":{"view_html":"https://pith.science/pith/L754SDXDCQQMYF35OIRBMTEVCZ","download_json":"https://pith.science/pith/L754SDXDCQQMYF35OIRBMTEVCZ.json","view_paper":"https://pith.science/paper/L754SDXD","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2601.09624&json=true","fetch_graph":"https://pith.science/api/pith-number/L754SDXDCQQMYF35OIRBMTEVCZ/graph.json","fetch_events":"https://pith.science/api/pith-number/L754SDXDCQQMYF35OIRBMTEVCZ/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/L754SDXDCQQMYF35OIRBMTEVCZ/action/timestamp_anchor","attest_storage":"https://pith.science/pith/L754SDXDCQQMYF35OIRBMTEVCZ/action/storage_attestation","attest_author":"https://pith.science/pith/L754SDXDCQQMYF35OIRBMTEVCZ/action/author_attestation","sign_citation":"https://pith.science/pith/L754SDXDCQQMYF35OIRBMTEVCZ/action/citation_signature","submit_replication":"https://pith.science/pith/L754SDXDCQQMYF35OIRBMTEVCZ/action/replication_record"}},"created_at":"2026-07-28T01:23:28.349341+00:00","updated_at":"2026-07-28T01:23:28.349341+00:00"}