{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2023:FDWBSYFX5H2TU6J7VSBNHEARAP","short_pith_number":"pith:FDWBSYFX","schema_version":"1.0","canonical_sha256":"28ec1960b7e9f53a793fac82d3901103f17e414c1c3cb2fd935ece5ba651d5e3","source":{"kind":"arxiv","id":"2309.10105","version":2},"attestation_state":"computed","paper":{"title":"Understanding Catastrophic Forgetting in Language Models via Implicit Inference","license":"http://creativecommons.org/licenses/by/4.0/","headline":"","cross_cats":["cs.LG"],"primary_cat":"cs.CL","authors_text":"Aditi Raghunathan, Jacob Mitchell Springer, Suhas Kotha","submitted_at":"2023-09-18T19:28:48Z","abstract_excerpt":"We lack a systematic understanding of the effects of fine-tuning (via methods such as instruction-tuning or reinforcement learning from human feedback), particularly on tasks outside the narrow fine-tuning distribution. In a simplified scenario, we demonstrate that improving performance on tasks within the fine-tuning data distribution comes at the expense of capabilities on other tasks. We hypothesize that language models implicitly infer the task of the prompt and that fine-tuning skews this inference towards tasks in the fine-tuning distribution. To test this, we propose Conjugate Prompting"},"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":"2309.10105","kind":"arxiv","version":2},"metadata":{"license":"http://creativecommons.org/licenses/by/4.0/","primary_cat":"cs.CL","submitted_at":"2023-09-18T19:28:48Z","cross_cats_sorted":["cs.LG"],"title_canon_sha256":"d3dbb70aae54e9f99fd252455f3a18c12fb77b0951b50a17c5d764bb2e08ee1b","abstract_canon_sha256":"9ccfef5b6d556b41f54777394e0fc8f961fef9d41210796dc578aeb56bf2cd96"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-07-05T08:07:29.346135Z","signature_b64":"TNUIbZPrPfdD/Y6KbSmvJOWl3aV1aDU+kicVKBEWl1OruvpRgX8L+Hx/D9NYxgwcZ5XFEiqCDPpYi1wObQqWBQ==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"28ec1960b7e9f53a793fac82d3901103f17e414c1c3cb2fd935ece5ba651d5e3","last_reissued_at":"2026-07-05T08:07:29.345590Z","signature_status":"signed_v1","first_computed_at":"2026-07-05T08:07:29.345590Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"Understanding Catastrophic Forgetting in Language Models via Implicit Inference","license":"http://creativecommons.org/licenses/by/4.0/","headline":"","cross_cats":["cs.LG"],"primary_cat":"cs.CL","authors_text":"Aditi Raghunathan, Jacob Mitchell Springer, Suhas Kotha","submitted_at":"2023-09-18T19:28:48Z","abstract_excerpt":"We lack a systematic understanding of the effects of fine-tuning (via methods such as instruction-tuning or reinforcement learning from human feedback), particularly on tasks outside the narrow fine-tuning distribution. In a simplified scenario, we demonstrate that improving performance on tasks within the fine-tuning data distribution comes at the expense of capabilities on other tasks. We hypothesize that language models implicitly infer the task of the prompt and that fine-tuning skews this inference towards tasks in the fine-tuning distribution. To test this, we propose Conjugate Prompting"},"claims":{"count":0,"items":[],"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"source":{"id":"2309.10105","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/2309.10105/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":"2309.10105","created_at":"2026-07-05T08:07:29.345650+00:00"},{"alias_kind":"arxiv_version","alias_value":"2309.10105v2","created_at":"2026-07-05T08:07:29.345650+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2309.10105","created_at":"2026-07-05T08:07:29.345650+00:00"},{"alias_kind":"pith_short_12","alias_value":"FDWBSYFX5H2T","created_at":"2026-07-05T08:07:29.345650+00:00"},{"alias_kind":"pith_short_16","alias_value":"FDWBSYFX5H2TU6J7","created_at":"2026-07-05T08:07:29.345650+00:00"},{"alias_kind":"pith_short_8","alias_value":"FDWBSYFX","created_at":"2026-07-05T08:07:29.345650+00:00"}],"events":[],"event_summary":{},"paper_claims":[],"inbound_citations":{"count":12,"internal_anchor_count":0,"sample":[{"citing_arxiv_id":"2606.18738","citing_title":"GRIDEX: Grid-Grounded Forensic Explanations for Deepfake Spectrogram Analysis","ref_index":37,"is_internal_anchor":false},{"citing_arxiv_id":"2606.03102","citing_title":"Small RL Controller, Large Language Model: RL-Guided Adaptive Sampling for Test-Time Scaling","ref_index":38,"is_internal_anchor":false},{"citing_arxiv_id":"2606.26102","citing_title":"Helpfulness Hurts: Domain-Dependent Degradation of Mid-Trained Compassion Values Under Post-Training","ref_index":12,"is_internal_anchor":false},{"citing_arxiv_id":"2606.28998","citing_title":"Reward-Free Code Alignment from Pretrained or Fine-Tuned LLM: Unpacking the Trade-offs for Code Generation","ref_index":26,"is_internal_anchor":false},{"citing_arxiv_id":"2606.29184","citing_title":"BaRA: Bayesian Adaptive Rank Allocation for Parameter-Efficient Fine-Tuning","ref_index":28,"is_internal_anchor":false},{"citing_arxiv_id":"2606.00147","citing_title":"RAFT: Data Refinement and Adaptive Distillation for Domain Fine-Tuning with Alleviated Forgetting","ref_index":30,"is_internal_anchor":false},{"citing_arxiv_id":"2603.06610","citing_title":"CapTrack: Multifaceted Evaluation of Forgetting in LLM Post-Training","ref_index":22,"is_internal_anchor":false},{"citing_arxiv_id":"2601.21577","citing_title":"Collaborative Parameter Learning: Mitigating Forgetting via Parameter-Level Gradient Analysis","ref_index":10,"is_internal_anchor":false},{"citing_arxiv_id":"2602.08813","citing_title":"Robust Policy Optimization to Prevent Catastrophic Forgetting","ref_index":30,"is_internal_anchor":false},{"citing_arxiv_id":"2605.11716","citing_title":"SafeSteer: A Decoding-level Defense Mechanism for Multimodal Large Language Models","ref_index":31,"is_internal_anchor":false},{"citing_arxiv_id":"2605.10120","citing_title":"MicroWorld: Empowering Multimodal Large Language Models to Bridge the Microscopic Domain Gap with Multimodal Attribute Graph","ref_index":13,"is_internal_anchor":false},{"citing_arxiv_id":"2604.12599","citing_title":"Beyond Pre-Training: The Full Lifecycle of Foundation Models on HPC Systems","ref_index":13,"is_internal_anchor":false}]},"formal_canon":{"evidence_count":0,"sample":[],"anchors":[]},"links":{"html":"https://pith.science/pith/FDWBSYFX5H2TU6J7VSBNHEARAP","json":"https://pith.science/pith/FDWBSYFX5H2TU6J7VSBNHEARAP.json","graph_json":"https://pith.science/api/pith-number/FDWBSYFX5H2TU6J7VSBNHEARAP/graph.json","events_json":"https://pith.science/api/pith-number/FDWBSYFX5H2TU6J7VSBNHEARAP/events.json","paper":"https://pith.science/paper/FDWBSYFX"},"agent_actions":{"view_html":"https://pith.science/pith/FDWBSYFX5H2TU6J7VSBNHEARAP","download_json":"https://pith.science/pith/FDWBSYFX5H2TU6J7VSBNHEARAP.json","view_paper":"https://pith.science/paper/FDWBSYFX","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2309.10105&json=true","fetch_graph":"https://pith.science/api/pith-number/FDWBSYFX5H2TU6J7VSBNHEARAP/graph.json","fetch_events":"https://pith.science/api/pith-number/FDWBSYFX5H2TU6J7VSBNHEARAP/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/FDWBSYFX5H2TU6J7VSBNHEARAP/action/timestamp_anchor","attest_storage":"https://pith.science/pith/FDWBSYFX5H2TU6J7VSBNHEARAP/action/storage_attestation","attest_author":"https://pith.science/pith/FDWBSYFX5H2TU6J7VSBNHEARAP/action/author_attestation","sign_citation":"https://pith.science/pith/FDWBSYFX5H2TU6J7VSBNHEARAP/action/citation_signature","submit_replication":"https://pith.science/pith/FDWBSYFX5H2TU6J7VSBNHEARAP/action/replication_record"}},"created_at":"2026-07-05T08:07:29.345650+00:00","updated_at":"2026-07-05T08:07:29.345650+00:00"}