{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2023:NXVTYTKASEU64VRAURF5FGLZE3","short_pith_number":"pith:NXVTYTKA","schema_version":"1.0","canonical_sha256":"6deb3c4d409129ee5620a44bd2997926d193548245832d6d42e88680ebc29e3f","source":{"kind":"arxiv","id":"2310.15213","version":2},"attestation_state":"computed","paper":{"title":"Function Vectors in Large Language Models","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","headline":"","cross_cats":["cs.LG"],"primary_cat":"cs.CL","authors_text":"Aaron Mueller, Arnab Sen Sharma, Byron C. Wallace, David Bau, Eric Todd, Millicent L. Li","submitted_at":"2023-10-23T17:55:24Z","abstract_excerpt":"We report the presence of a simple neural mechanism that represents an input-output function as a vector within autoregressive transformer language models (LMs). Using causal mediation analysis on a diverse range of in-context-learning (ICL) tasks, we find that a small number attention heads transport a compact representation of the demonstrated task, which we call a function vector (FV). FVs are robust to changes in context, i.e., they trigger execution of the task on inputs such as zero-shot and natural text settings that do not resemble the ICL contexts from which they are collected. We tes"},"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":"2310.15213","kind":"arxiv","version":2},"metadata":{"license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","primary_cat":"cs.CL","submitted_at":"2023-10-23T17:55:24Z","cross_cats_sorted":["cs.LG"],"title_canon_sha256":"8f48fd39e9fe1b8330769574510676a896ad3446c270e96186425df15daf5314","abstract_canon_sha256":"2287ea84179175b1ea860881a818c70f83605a80cd29d62cbfdf57c3977b643c"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-07-05T07:48:49.557484Z","signature_b64":"ozBBEewfUfWYGUT5xw9yo3cPAHZNDuoqMsi1wydi+xhV5WsEXWG9g/RgkYw8hrvZg7guLGvXxHqxduXGzBz5Cw==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"6deb3c4d409129ee5620a44bd2997926d193548245832d6d42e88680ebc29e3f","last_reissued_at":"2026-07-05T07:48:49.557045Z","signature_status":"signed_v1","first_computed_at":"2026-07-05T07:48:49.557045Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"Function Vectors in Large Language Models","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","headline":"","cross_cats":["cs.LG"],"primary_cat":"cs.CL","authors_text":"Aaron Mueller, Arnab Sen Sharma, Byron C. Wallace, David Bau, Eric Todd, Millicent L. Li","submitted_at":"2023-10-23T17:55:24Z","abstract_excerpt":"We report the presence of a simple neural mechanism that represents an input-output function as a vector within autoregressive transformer language models (LMs). Using causal mediation analysis on a diverse range of in-context-learning (ICL) tasks, we find that a small number attention heads transport a compact representation of the demonstrated task, which we call a function vector (FV). FVs are robust to changes in context, i.e., they trigger execution of the task on inputs such as zero-shot and natural text settings that do not resemble the ICL contexts from which they are collected. We tes"},"claims":{"count":0,"items":[],"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"source":{"id":"2310.15213","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/2310.15213/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":"2310.15213","created_at":"2026-07-05T07:48:49.557097+00:00"},{"alias_kind":"arxiv_version","alias_value":"2310.15213v2","created_at":"2026-07-05T07:48:49.557097+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2310.15213","created_at":"2026-07-05T07:48:49.557097+00:00"},{"alias_kind":"pith_short_12","alias_value":"NXVTYTKASEU6","created_at":"2026-07-05T07:48:49.557097+00:00"},{"alias_kind":"pith_short_16","alias_value":"NXVTYTKASEU64VRA","created_at":"2026-07-05T07:48:49.557097+00:00"},{"alias_kind":"pith_short_8","alias_value":"NXVTYTKA","created_at":"2026-07-05T07:48:49.557097+00:00"}],"events":[],"event_summary":{},"paper_claims":[],"inbound_citations":{"count":26,"internal_anchor_count":0,"sample":[{"citing_arxiv_id":"2606.26660","citing_title":"Generating Special Triangulations with Transformers","ref_index":49,"is_internal_anchor":false},{"citing_arxiv_id":"2606.24937","citing_title":"The Hitchhiker's Guide to Agentic AI: From Foundations to Systems","ref_index":128,"is_internal_anchor":false},{"citing_arxiv_id":"2606.10929","citing_title":"Recoverable but Not Stationary:Local Linear Structures in Weights and Activations","ref_index":18,"is_internal_anchor":false},{"citing_arxiv_id":"2606.04434","citing_title":"Hyper-ICL: Attention Calibration with Hyperbolic Anchor Distillation for Multimodal In-Context Learning","ref_index":9,"is_internal_anchor":false},{"citing_arxiv_id":"2605.20382","citing_title":"Do as I Say, Not as I Do: Instruction-Induction Conflict in LLMs","ref_index":20,"is_internal_anchor":false},{"citing_arxiv_id":"2605.25225","citing_title":"Transformer Field Theory: A Response-Theoretic Approach to Mechanistic Interpretability","ref_index":21,"is_internal_anchor":false},{"citing_arxiv_id":"2605.29358","citing_title":"Scaling Monosemanticity: Extracting Interpretable Features from Claude 3 Sonnet","ref_index":71,"is_internal_anchor":false},{"citing_arxiv_id":"2605.23180","citing_title":"Self-Improving In-Context Learning","ref_index":45,"is_internal_anchor":false},{"citing_arxiv_id":"2605.20382","citing_title":"Do as I Say, Not as I Do: Instruction-Induction Conflict in LLMs","ref_index":20,"is_internal_anchor":false},{"citing_arxiv_id":"2605.15053","citing_title":"TFGN: Task-Free, Replay-Free Continual Pre-Training Without Catastrophic Forgetting at LLM Scale","ref_index":67,"is_internal_anchor":false},{"citing_arxiv_id":"2507.20906","citing_title":"Soft Head Selection for Injecting ICL-Derived Task Embeddings","ref_index":20,"is_internal_anchor":false},{"citing_arxiv_id":"2509.24164","citing_title":"Localizing Task Recognition and Task Learning in In-Context Learning via Attention Head Analysis","ref_index":41,"is_internal_anchor":false},{"citing_arxiv_id":"2510.01685","citing_title":"How Do Language Models Compose Functions?","ref_index":38,"is_internal_anchor":false},{"citing_arxiv_id":"2404.15255","citing_title":"How to use and interpret activation patching","ref_index":27,"is_internal_anchor":false},{"citing_arxiv_id":"2605.12412","citing_title":"Stories in Space: In-Context Learning Trajectories in Conceptual Belief Space","ref_index":103,"is_internal_anchor":false},{"citing_arxiv_id":"2605.06225","citing_title":"Memory Inception: Latent-Space KV Cache Manipulation for Steering LLMs","ref_index":21,"is_internal_anchor":false},{"citing_arxiv_id":"2311.03658","citing_title":"The Linear Representation Hypothesis and the Geometry of Large Language Models","ref_index":23,"is_internal_anchor":false},{"citing_arxiv_id":"2604.23371","citing_title":"When Context Sticks: Studying Interference in In-Context Learning","ref_index":22,"is_internal_anchor":false},{"citing_arxiv_id":"2605.05715","citing_title":"Decodable but Not Corrected by Fixed Residual-Stream Linear Steering: Evidence from Medical LLM Failure Regimes","ref_index":62,"is_internal_anchor":false},{"citing_arxiv_id":"2605.06225","citing_title":"Memory Inception: Latent-Space KV Cache Manipulation for Steering LLMs","ref_index":21,"is_internal_anchor":false},{"citing_arxiv_id":"2605.05115","citing_title":"Manifold Steering Reveals the Shared Geometry of Neural Network Representation and Behavior","ref_index":232,"is_internal_anchor":false},{"citing_arxiv_id":"2604.10326","citing_title":"Jailbreaking the Matrix: Nullspace Steering for Controlled Model Subversion","ref_index":43,"is_internal_anchor":false},{"citing_arxiv_id":"2605.07990","citing_title":"Tool Calling is Linearly Readable and Steerable in Language Models","ref_index":71,"is_internal_anchor":false},{"citing_arxiv_id":"2604.06377","citing_title":"The Master Key Hypothesis: Unlocking Cross-Model Capability Transfer via Linear Subspace Alignment","ref_index":65,"is_internal_anchor":false},{"citing_arxiv_id":"2604.17663","citing_title":"ATLAS: Constitution-Conditioned Latent Geometry and Redistribution Across Language Models and Neural Perturbation Data","ref_index":16,"is_internal_anchor":false}]},"formal_canon":{"evidence_count":0,"sample":[],"anchors":[]},"links":{"html":"https://pith.science/pith/NXVTYTKASEU64VRAURF5FGLZE3","json":"https://pith.science/pith/NXVTYTKASEU64VRAURF5FGLZE3.json","graph_json":"https://pith.science/api/pith-number/NXVTYTKASEU64VRAURF5FGLZE3/graph.json","events_json":"https://pith.science/api/pith-number/NXVTYTKASEU64VRAURF5FGLZE3/events.json","paper":"https://pith.science/paper/NXVTYTKA"},"agent_actions":{"view_html":"https://pith.science/pith/NXVTYTKASEU64VRAURF5FGLZE3","download_json":"https://pith.science/pith/NXVTYTKASEU64VRAURF5FGLZE3.json","view_paper":"https://pith.science/paper/NXVTYTKA","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2310.15213&json=true","fetch_graph":"https://pith.science/api/pith-number/NXVTYTKASEU64VRAURF5FGLZE3/graph.json","fetch_events":"https://pith.science/api/pith-number/NXVTYTKASEU64VRAURF5FGLZE3/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/NXVTYTKASEU64VRAURF5FGLZE3/action/timestamp_anchor","attest_storage":"https://pith.science/pith/NXVTYTKASEU64VRAURF5FGLZE3/action/storage_attestation","attest_author":"https://pith.science/pith/NXVTYTKASEU64VRAURF5FGLZE3/action/author_attestation","sign_citation":"https://pith.science/pith/NXVTYTKASEU64VRAURF5FGLZE3/action/citation_signature","submit_replication":"https://pith.science/pith/NXVTYTKASEU64VRAURF5FGLZE3/action/replication_record"}},"created_at":"2026-07-05T07:48:49.557097+00:00","updated_at":"2026-07-05T07:48:49.557097+00:00"}