{"paper":{"title":"Less is More: Geometric Unlearning for LLMs with Minimal Data Disclosure","license":"http://creativecommons.org/licenses/by/4.0/","headline":"Geometric Unlearning lets LLMs forget specific private facts using only a handful of synthetic prompts while retaining general performance.","cross_cats":[],"primary_cat":"cs.CL","authors_text":"Chenchen Tan, Cunjian Chen, Longxiang Gao, Shujie Cui, Xinghao Li, Youyang Qu","submitted_at":"2026-05-03T06:20:03Z","abstract_excerpt":"As large language models (LLMs) are increasingly deployed in real-world systems, they must support post-hoc removal of specific content to meet privacy and governance requirements. This motivates selective unlearning, which suppresses information about a particular entity or topic while preserving the LLM's general utility. However, most existing LLM unlearning methods require access to the original training corpus and rely on output-level refusal tuning or broad gradient updates, creating a tension among unlearning strength, non-target preservation, and data availability. We propose Geometric"},"claims":{"count":4,"items":[{"kind":"strongest_claim","text":"Across privacy-oriented unlearning benchmarks (ToFU and UnlearnPII), GU achieves strong target suppression with minimal impact on non-target performance, demonstrating that effective unlearning can be achieved with minimal synthetic data.","source":"verdict.strongest_claim","status":"machine_extracted","claim_id":"C1","attestation":"unclaimed"},{"kind":"weakest_assumption","text":"The assumption that a compact low-rank geometry distilled from safe reference prompts, combined with projection-based alignment using synthetic anchors, can effectively suppress target information in the model's hidden states without access to the original training corpus or significant utility loss.","source":"verdict.weakest_assumption","status":"machine_extracted","claim_id":"C2","attestation":"unclaimed"},{"kind":"one_line_summary","text":"Geometric Unlearning suppresses specific knowledge in LLMs by projecting hidden planning states onto a low-rank safe geometry derived from minimal reference prompts.","source":"verdict.one_line_summary","status":"machine_extracted","claim_id":"C3","attestation":"unclaimed"},{"kind":"headline","text":"Geometric Unlearning lets LLMs forget specific private facts using only a handful of synthetic prompts while retaining general performance.","source":"verdict.pith_extraction.headline","status":"machine_extracted","claim_id":"C4","attestation":"unclaimed"}],"snapshot_sha256":"60b0f8712d9d23ca96357302aab8f3362d99a638b9a2f4d51e0250f8b44530c0"},"source":{"id":"2605.01735","kind":"arxiv","version":2},"verdict":{"id":"504d7cdf-3774-41e3-b5cf-790afc4c786a","model_set":{"reader":"grok-4.3"},"created_at":"2026-05-10T15:58:11.548449Z","strongest_claim":"Across privacy-oriented unlearning benchmarks (ToFU and UnlearnPII), GU achieves strong target suppression with minimal impact on non-target performance, demonstrating that effective unlearning can be achieved with minimal synthetic data.","one_line_summary":"Geometric Unlearning suppresses specific knowledge in LLMs by projecting hidden planning states onto a low-rank safe geometry derived from minimal reference prompts.","pipeline_version":"pith-pipeline@v0.9.0","weakest_assumption":"The assumption that a compact low-rank geometry distilled from safe reference prompts, combined with projection-based alignment using synthetic anchors, can effectively suppress target information in the model's hidden states without access to the original training corpus or significant utility loss.","pith_extraction_headline":"Geometric Unlearning lets LLMs forget specific private facts using only a handful of synthetic prompts while retaining general performance."},"integrity":{"clean":true,"summary":{"advisory":0,"critical":0,"by_detector":{},"informational":0},"endpoint":"/pith/2605.01735/integrity.json","findings":[],"available":true,"detectors_run":[{"name":"ai_meta_artifact","ran_at":"2026-05-20T17:37:19.650148Z","status":"completed","version":"1.0.0","findings_count":0},{"name":"doi_title_agreement","ran_at":"2026-05-20T05:01:23.091965Z","status":"completed","version":"1.0.0","findings_count":0},{"name":"doi_compliance","ran_at":"2026-05-19T17:00:50.088978Z","status":"completed","version":"1.0.0","findings_count":0}],"snapshot_sha256":"2792726b88a81447be03310858e3b22ff01e6d40439872c768522f65e9d3fe66"},"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"}