{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2025:AHK6R6L4KX26YCRPTLXS2MCEO3","short_pith_number":"pith:AHK6R6L4","schema_version":"1.0","canonical_sha256":"01d5e8f97c55f5ec0a2f9aef2d304476c0cbd3f6d8ea01977f5cea4f70fda3f8","source":{"kind":"arxiv","id":"2508.02079","version":1},"attestation_state":"computed","paper":{"title":"AlignGuard-LoRA: Alignment-Preserving Fine-Tuning via Fisher-Guided Decomposition and Riemannian-Geodesic Collision Regularization","license":"http://creativecommons.org/licenses/by/4.0/","headline":"","cross_cats":["cs.AI"],"primary_cat":"cs.LG","authors_text":"Abhilekh Borah, Aman Chadha, Amitava Das, Vinija Jain","submitted_at":"2025-08-04T05:45:24Z","abstract_excerpt":"Low-rank adaptation (LoRA) has become a standard tool for efficiently fine-tuning large language models (LLMs). Yet, even minor LoRA updates can induce alignment drift, weakening safety and behavioral constraints through entangled parameter changes. To address this, we propose AlignGuard-LoRA (AGL), a principled framework for preserving alignment during finetuning. AGL introduces several key components: a primary task loss for supervision, Fisher Information Matrix-based regularization to restrict updates in alignment-sensitive subspaces, and task-specific regularization to stabilize the integ"},"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":"2508.02079","kind":"arxiv","version":1},"metadata":{"license":"http://creativecommons.org/licenses/by/4.0/","primary_cat":"cs.LG","submitted_at":"2025-08-04T05:45:24Z","cross_cats_sorted":["cs.AI"],"title_canon_sha256":"7a0b4a5e76282d18e8e3957d7869148042254f9e46dd3ee276349f66bfd82dbb","abstract_canon_sha256":"493b0a630bfd050a81a253537104e889aaf9cd31d4d8e4a2be1ff445658cb130"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-07-05T11:47:50.539608Z","signature_b64":"J+Rlp+0SWCGzlXDOzgm6IO9cBWow1/6bk9JnnN74lp3I7n/U+jQmk5C7Q+S4a4rIsFFmF8iFSGogfI+Xlzb5AQ==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"01d5e8f97c55f5ec0a2f9aef2d304476c0cbd3f6d8ea01977f5cea4f70fda3f8","last_reissued_at":"2026-07-05T11:47:50.539182Z","signature_status":"signed_v1","first_computed_at":"2026-07-05T11:47:50.539182Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"AlignGuard-LoRA: Alignment-Preserving Fine-Tuning via Fisher-Guided Decomposition and Riemannian-Geodesic Collision Regularization","license":"http://creativecommons.org/licenses/by/4.0/","headline":"","cross_cats":["cs.AI"],"primary_cat":"cs.LG","authors_text":"Abhilekh Borah, Aman Chadha, Amitava Das, Vinija Jain","submitted_at":"2025-08-04T05:45:24Z","abstract_excerpt":"Low-rank adaptation (LoRA) has become a standard tool for efficiently fine-tuning large language models (LLMs). Yet, even minor LoRA updates can induce alignment drift, weakening safety and behavioral constraints through entangled parameter changes. To address this, we propose AlignGuard-LoRA (AGL), a principled framework for preserving alignment during finetuning. AGL introduces several key components: a primary task loss for supervision, Fisher Information Matrix-based regularization to restrict updates in alignment-sensitive subspaces, and task-specific regularization to stabilize the integ"},"claims":{"count":0,"items":[],"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"source":{"id":"2508.02079","kind":"arxiv","version":1},"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/2508.02079/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":"2508.02079","created_at":"2026-07-05T11:47:50.539232+00:00"},{"alias_kind":"arxiv_version","alias_value":"2508.02079v1","created_at":"2026-07-05T11:47:50.539232+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2508.02079","created_at":"2026-07-05T11:47:50.539232+00:00"},{"alias_kind":"pith_short_12","alias_value":"AHK6R6L4KX26","created_at":"2026-07-05T11:47:50.539232+00:00"},{"alias_kind":"pith_short_16","alias_value":"AHK6R6L4KX26YCRP","created_at":"2026-07-05T11:47:50.539232+00:00"},{"alias_kind":"pith_short_8","alias_value":"AHK6R6L4","created_at":"2026-07-05T11:47:50.539232+00:00"}],"events":[],"event_summary":{},"paper_claims":[],"inbound_citations":{"count":2,"internal_anchor_count":0,"sample":[{"citing_arxiv_id":"2606.15980","citing_title":"Do Activation Monitors Survive Model Updates? Benchmarking, Predicting, and Repairing Activation-Monitor Staleness","ref_index":44,"is_internal_anchor":false},{"citing_arxiv_id":"2605.01913","citing_title":"RefusalGuard: Geometry-Preserving Fine-Tuning for Safety in LLMs","ref_index":2,"is_internal_anchor":false}]},"formal_canon":{"evidence_count":0,"sample":[],"anchors":[]},"links":{"html":"https://pith.science/pith/AHK6R6L4KX26YCRPTLXS2MCEO3","json":"https://pith.science/pith/AHK6R6L4KX26YCRPTLXS2MCEO3.json","graph_json":"https://pith.science/api/pith-number/AHK6R6L4KX26YCRPTLXS2MCEO3/graph.json","events_json":"https://pith.science/api/pith-number/AHK6R6L4KX26YCRPTLXS2MCEO3/events.json","paper":"https://pith.science/paper/AHK6R6L4"},"agent_actions":{"view_html":"https://pith.science/pith/AHK6R6L4KX26YCRPTLXS2MCEO3","download_json":"https://pith.science/pith/AHK6R6L4KX26YCRPTLXS2MCEO3.json","view_paper":"https://pith.science/paper/AHK6R6L4","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2508.02079&json=true","fetch_graph":"https://pith.science/api/pith-number/AHK6R6L4KX26YCRPTLXS2MCEO3/graph.json","fetch_events":"https://pith.science/api/pith-number/AHK6R6L4KX26YCRPTLXS2MCEO3/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/AHK6R6L4KX26YCRPTLXS2MCEO3/action/timestamp_anchor","attest_storage":"https://pith.science/pith/AHK6R6L4KX26YCRPTLXS2MCEO3/action/storage_attestation","attest_author":"https://pith.science/pith/AHK6R6L4KX26YCRPTLXS2MCEO3/action/author_attestation","sign_citation":"https://pith.science/pith/AHK6R6L4KX26YCRPTLXS2MCEO3/action/citation_signature","submit_replication":"https://pith.science/pith/AHK6R6L4KX26YCRPTLXS2MCEO3/action/replication_record"}},"created_at":"2026-07-05T11:47:50.539232+00:00","updated_at":"2026-07-05T11:47:50.539232+00:00"}