{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2025:GGIKJM6L7PZ67OTDAIKAREOUWX","short_pith_number":"pith:GGIKJM6L","schema_version":"1.0","canonical_sha256":"3190a4b3cbfbf3efba6302140891d4b5fac56ac8e0412d6db1ad15557c1c6ba7","source":{"kind":"arxiv","id":"2503.20807","version":1},"attestation_state":"computed","paper":{"title":"Fundamental Safety-Capability Trade-offs in Fine-tuning Large Language Models","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","headline":"","cross_cats":["cs.AI","cs.CL","cs.LG"],"primary_cat":"stat.ML","authors_text":"Han Shen, Payel Das, Pin-Yu Chen, Tianyi Chen","submitted_at":"2025-03-24T20:41:57Z","abstract_excerpt":"Fine-tuning Large Language Models (LLMs) on some task-specific datasets has been a primary use of LLMs. However, it has been empirically observed that this approach to enhancing capability inevitably compromises safety, a phenomenon also known as the safety-capability trade-off in LLM fine-tuning. This paper presents a theoretical framework for understanding the interplay between safety and capability in two primary safety-aware LLM fine-tuning strategies, providing new insights into the effects of data similarity, context overlap, and alignment loss landscape. Our theoretical results characte"},"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":"2503.20807","kind":"arxiv","version":1},"metadata":{"license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","primary_cat":"stat.ML","submitted_at":"2025-03-24T20:41:57Z","cross_cats_sorted":["cs.AI","cs.CL","cs.LG"],"title_canon_sha256":"8dc2cc9bba9912be6a21491d6f668073cad27bf203b017a2380604817ab3dbb2","abstract_canon_sha256":"3afed9d3d891c92c620f82f4c828fccb1132f69903699cbd8be88f57bd9e7bfa"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-07-05T10:39:48.553186Z","signature_b64":"3oa1cPQsrYyqfgKRyPzqX76Jw0pwNsGy/KSYrKskqLjXsXoF5a6ftZeR9Up6mLWosPEpr1Hu3x9IbLLC6a7eCg==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"3190a4b3cbfbf3efba6302140891d4b5fac56ac8e0412d6db1ad15557c1c6ba7","last_reissued_at":"2026-07-05T10:39:48.552695Z","signature_status":"signed_v1","first_computed_at":"2026-07-05T10:39:48.552695Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"Fundamental Safety-Capability Trade-offs in Fine-tuning Large Language Models","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","headline":"","cross_cats":["cs.AI","cs.CL","cs.LG"],"primary_cat":"stat.ML","authors_text":"Han Shen, Payel Das, Pin-Yu Chen, Tianyi Chen","submitted_at":"2025-03-24T20:41:57Z","abstract_excerpt":"Fine-tuning Large Language Models (LLMs) on some task-specific datasets has been a primary use of LLMs. However, it has been empirically observed that this approach to enhancing capability inevitably compromises safety, a phenomenon also known as the safety-capability trade-off in LLM fine-tuning. This paper presents a theoretical framework for understanding the interplay between safety and capability in two primary safety-aware LLM fine-tuning strategies, providing new insights into the effects of data similarity, context overlap, and alignment loss landscape. Our theoretical results characte"},"claims":{"count":0,"items":[],"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"source":{"id":"2503.20807","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/2503.20807/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":"2503.20807","created_at":"2026-07-05T10:39:48.552747+00:00"},{"alias_kind":"arxiv_version","alias_value":"2503.20807v1","created_at":"2026-07-05T10:39:48.552747+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2503.20807","created_at":"2026-07-05T10:39:48.552747+00:00"},{"alias_kind":"pith_short_12","alias_value":"GGIKJM6L7PZ6","created_at":"2026-07-05T10:39:48.552747+00:00"},{"alias_kind":"pith_short_16","alias_value":"GGIKJM6L7PZ67OTD","created_at":"2026-07-05T10:39:48.552747+00:00"},{"alias_kind":"pith_short_8","alias_value":"GGIKJM6L","created_at":"2026-07-05T10:39:48.552747+00:00"}],"events":[],"event_summary":{},"paper_claims":[],"inbound_citations":{"count":4,"internal_anchor_count":0,"sample":[{"citing_arxiv_id":"2606.26102","citing_title":"Helpfulness Hurts: Domain-Dependent Degradation of Mid-Trained Compassion Values Under Post-Training","ref_index":6,"is_internal_anchor":false},{"citing_arxiv_id":"2605.23024","citing_title":"The Deterministic Horizon: Impossibility Results as Design Specifications for Trustworthy AI Systems","ref_index":128,"is_internal_anchor":false},{"citing_arxiv_id":"2602.17546","citing_title":"Learning to Stay Safe: Adaptive Regularization Against Safety Degradation during Fine-Tuning","ref_index":1,"is_internal_anchor":false},{"citing_arxiv_id":"2605.04572","citing_title":"From Parameter Dynamics to Risk Scoring : Quantifying Sample-Level Safety Degradation in LLM Fine-tuning","ref_index":3,"is_internal_anchor":false}]},"formal_canon":{"evidence_count":0,"sample":[],"anchors":[]},"links":{"html":"https://pith.science/pith/GGIKJM6L7PZ67OTDAIKAREOUWX","json":"https://pith.science/pith/GGIKJM6L7PZ67OTDAIKAREOUWX.json","graph_json":"https://pith.science/api/pith-number/GGIKJM6L7PZ67OTDAIKAREOUWX/graph.json","events_json":"https://pith.science/api/pith-number/GGIKJM6L7PZ67OTDAIKAREOUWX/events.json","paper":"https://pith.science/paper/GGIKJM6L"},"agent_actions":{"view_html":"https://pith.science/pith/GGIKJM6L7PZ67OTDAIKAREOUWX","download_json":"https://pith.science/pith/GGIKJM6L7PZ67OTDAIKAREOUWX.json","view_paper":"https://pith.science/paper/GGIKJM6L","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2503.20807&json=true","fetch_graph":"https://pith.science/api/pith-number/GGIKJM6L7PZ67OTDAIKAREOUWX/graph.json","fetch_events":"https://pith.science/api/pith-number/GGIKJM6L7PZ67OTDAIKAREOUWX/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/GGIKJM6L7PZ67OTDAIKAREOUWX/action/timestamp_anchor","attest_storage":"https://pith.science/pith/GGIKJM6L7PZ67OTDAIKAREOUWX/action/storage_attestation","attest_author":"https://pith.science/pith/GGIKJM6L7PZ67OTDAIKAREOUWX/action/author_attestation","sign_citation":"https://pith.science/pith/GGIKJM6L7PZ67OTDAIKAREOUWX/action/citation_signature","submit_replication":"https://pith.science/pith/GGIKJM6L7PZ67OTDAIKAREOUWX/action/replication_record"}},"created_at":"2026-07-05T10:39:48.552747+00:00","updated_at":"2026-07-05T10:39:48.552747+00:00"}