{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2026:GUD5X3KTZYOHJ3YCGCV4XF3LGQ","short_pith_number":"pith:GUD5X3KT","schema_version":"1.0","canonical_sha256":"3507dbed53ce1c74ef0230abcb976b343c0a43b6b181873b53bc79e4fbdab568","source":{"kind":"arxiv","id":"2605.12765","version":1},"attestation_state":"computed","paper":{"title":"Inference-Time Machine Unlearning via Gated Activation Redirection","license":"http://creativecommons.org/licenses/by/4.0/","headline":"A gated input-dependent rotation in the residual stream enables inference-time unlearning of specific data in LLMs without altering weights.","cross_cats":[],"primary_cat":"cs.LG","authors_text":"Christian Mattjie, Flavio du Pin Calmon, Joana Pasquali, Jo\\~ao Vitor Boer Abitante, Kristen K. Arguello, Lucas S. Kupssinsk\\\"u, Ot\\'avio Parraga, Ramiro N. Barros, Rodrigo C. Barros, Vin\\'icius Conte Turani","submitted_at":"2026-05-12T21:26:25Z","abstract_excerpt":"Large Language Models memorize vast amounts of training data, raising concerns regarding privacy, copyright infringement, and safety. Machine unlearning seeks to remove the influence of a targeted forget set while preserving model performance, ideally approximating a model retrained from scratch without the forget set. Existing approaches aim to achieve this by updating model parameters via gradient-based methods. However, these updates are computationally expensive, lead to irreversible weight changes, and degrade when the model is quantized for deployment. A recent alternative to changing mo"},"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":"2605.12765","kind":"arxiv","version":1},"metadata":{"license":"http://creativecommons.org/licenses/by/4.0/","primary_cat":"cs.LG","submitted_at":"2026-05-12T21:26:25Z","cross_cats_sorted":[],"title_canon_sha256":"f7ac0148a6ac44443446671da58bd82b96960f87e1e7ee0de219e17cb0d846d7","abstract_canon_sha256":"98b122ffa4d5a003357308d804bbb6194227a3a3366125f7e6df692a29b1174a"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-05-18T03:09:48.304324Z","signature_b64":"I9KJTG/g9JMSLIHIGBYPafqocAzvbXnlHcsQhb16/SwTMkriIMHjtljDQjdOpnoPrXk80L+7HvN0xoRcAA/ECQ==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"3507dbed53ce1c74ef0230abcb976b343c0a43b6b181873b53bc79e4fbdab568","last_reissued_at":"2026-05-18T03:09:48.303388Z","signature_status":"signed_v1","first_computed_at":"2026-05-18T03:09:48.303388Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"Inference-Time Machine Unlearning via Gated Activation Redirection","license":"http://creativecommons.org/licenses/by/4.0/","headline":"A gated input-dependent rotation in the residual stream enables inference-time unlearning of specific data in LLMs without altering weights.","cross_cats":[],"primary_cat":"cs.LG","authors_text":"Christian Mattjie, Flavio du Pin Calmon, Joana Pasquali, Jo\\~ao Vitor Boer Abitante, Kristen K. Arguello, Lucas S. Kupssinsk\\\"u, Ot\\'avio Parraga, Ramiro N. Barros, Rodrigo C. Barros, Vin\\'icius Conte Turani","submitted_at":"2026-05-12T21:26:25Z","abstract_excerpt":"Large Language Models memorize vast amounts of training data, raising concerns regarding privacy, copyright infringement, and safety. Machine unlearning seeks to remove the influence of a targeted forget set while preserving model performance, ideally approximating a model retrained from scratch without the forget set. Existing approaches aim to achieve this by updating model parameters via gradient-based methods. However, these updates are computationally expensive, lead to irreversible weight changes, and degrade when the model is quantized for deployment. A recent alternative to changing mo"},"claims":{"count":4,"items":[{"kind":"strongest_claim","text":"GUARD-IT matches or exceeds 12 gradient-based baselines across three model scales, while being the only method to simultaneously preserve utility, suppress memorization, and avoid catastrophic collapse across all settings.","source":"verdict.strongest_claim","status":"machine_extracted","claim_id":"C1","attestation":"unclaimed"},{"kind":"weakest_assumption","text":"That an input-dependent norm-preserving rotation in the residual stream can selectively remove the influence of a forget set without introducing unintended changes to model behavior on unrelated inputs.","source":"verdict.weakest_assumption","status":"machine_extracted","claim_id":"C2","attestation":"unclaimed"},{"kind":"one_line_summary","text":"GUARD-IT performs machine unlearning in LLMs via inference-time gated activation redirection, matching or exceeding gradient-based baselines on TOFU and MUSE while preserving utility and working under quantization.","source":"verdict.one_line_summary","status":"machine_extracted","claim_id":"C3","attestation":"unclaimed"},{"kind":"headline","text":"A gated input-dependent rotation in the residual stream enables inference-time unlearning of specific data in LLMs without altering weights.","source":"verdict.pith_extraction.headline","status":"machine_extracted","claim_id":"C4","attestation":"unclaimed"}],"snapshot_sha256":"056e6c7803e400d74a30cce0f9bf9c96f56486ac3b14985c55b5492b53e5da86"},"source":{"id":"2605.12765","kind":"arxiv","version":1},"verdict":{"id":"9d004799-7886-4cb1-a9fd-78e8401a902b","model_set":{"reader":"grok-4.3"},"created_at":"2026-05-14T20:54:49.153797Z","strongest_claim":"GUARD-IT matches or exceeds 12 gradient-based baselines across three model scales, while being the only method to simultaneously preserve utility, suppress memorization, and avoid catastrophic collapse across all settings.","one_line_summary":"GUARD-IT performs machine unlearning in LLMs via inference-time gated activation redirection, matching or exceeding gradient-based baselines on TOFU and MUSE while preserving utility and working under quantization.","pipeline_version":"pith-pipeline@v0.9.0","weakest_assumption":"That an input-dependent norm-preserving rotation in the residual stream can selectively remove the influence of a forget set without introducing unintended changes to model behavior on unrelated inputs.","pith_extraction_headline":"A gated input-dependent rotation in the residual stream enables inference-time unlearning of specific data in LLMs without altering weights."},"references":{"count":59,"sample":[{"doi":"","year":null,"title":"TOFU: A Task of Fictitious Unlearning for LLMs","work_id":"b05148d7-e817-49f0-be96-3e297fddec9b","ref_index":1,"cited_arxiv_id":"2401.06121","is_internal_anchor":true},{"doi":"","year":null,"title":"Sentence-BERT: Sentence Embeddings using Siamese BERT-Networks , url =","work_id":"35a3edc1-58a2-4a50-b179-e2a6ccd771d2","ref_index":2,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2019,"title":"Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP). 2019","work_id":"e3609b2c-d523-4f5c-9a53-c3fd668e03ca","ref_index":3,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":null,"title":"Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , year =","work_id":"3296f847-5cec-41c6-a43b-fe9b45396798","ref_index":4,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2025,"title":"Smith and Chiyuan Zhang , booktitle=","work_id":"69af5d1c-1676-41d0-87e8-3d2cca8af021","ref_index":5,"cited_arxiv_id":"","is_internal_anchor":false}],"resolved_work":59,"snapshot_sha256":"41719a12ce8e1ff2064c05920c0a3b51fb65c535a3c9bd192bfaaab45e3370f5","internal_anchors":4},"formal_canon":{"evidence_count":2,"snapshot_sha256":"98b4bb989dfa4a6cf31c8e24baa4ad047ace9d455f8adfb46d21374b4c86be51"},"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":"2605.12765","created_at":"2026-05-18T03:09:48.303545+00:00"},{"alias_kind":"arxiv_version","alias_value":"2605.12765v1","created_at":"2026-05-18T03:09:48.303545+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2605.12765","created_at":"2026-05-18T03:09:48.303545+00:00"},{"alias_kind":"pith_short_12","alias_value":"GUD5X3KTZYOH","created_at":"2026-05-18T12:33:37.589309+00:00"},{"alias_kind":"pith_short_16","alias_value":"GUD5X3KTZYOHJ3YC","created_at":"2026-05-18T12:33:37.589309+00:00"},{"alias_kind":"pith_short_8","alias_value":"GUD5X3KT","created_at":"2026-05-18T12:33:37.589309+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/GUD5X3KTZYOHJ3YCGCV4XF3LGQ","json":"https://pith.science/pith/GUD5X3KTZYOHJ3YCGCV4XF3LGQ.json","graph_json":"https://pith.science/api/pith-number/GUD5X3KTZYOHJ3YCGCV4XF3LGQ/graph.json","events_json":"https://pith.science/api/pith-number/GUD5X3KTZYOHJ3YCGCV4XF3LGQ/events.json","paper":"https://pith.science/paper/GUD5X3KT"},"agent_actions":{"view_html":"https://pith.science/pith/GUD5X3KTZYOHJ3YCGCV4XF3LGQ","download_json":"https://pith.science/pith/GUD5X3KTZYOHJ3YCGCV4XF3LGQ.json","view_paper":"https://pith.science/paper/GUD5X3KT","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2605.12765&json=true","fetch_graph":"https://pith.science/api/pith-number/GUD5X3KTZYOHJ3YCGCV4XF3LGQ/graph.json","fetch_events":"https://pith.science/api/pith-number/GUD5X3KTZYOHJ3YCGCV4XF3LGQ/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/GUD5X3KTZYOHJ3YCGCV4XF3LGQ/action/timestamp_anchor","attest_storage":"https://pith.science/pith/GUD5X3KTZYOHJ3YCGCV4XF3LGQ/action/storage_attestation","attest_author":"https://pith.science/pith/GUD5X3KTZYOHJ3YCGCV4XF3LGQ/action/author_attestation","sign_citation":"https://pith.science/pith/GUD5X3KTZYOHJ3YCGCV4XF3LGQ/action/citation_signature","submit_replication":"https://pith.science/pith/GUD5X3KTZYOHJ3YCGCV4XF3LGQ/action/replication_record"}},"created_at":"2026-05-18T03:09:48.303545+00:00","updated_at":"2026-05-18T03:09:48.303545+00:00"}