{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2023:FVMKECSPVYLV34B5JLGYBO62JC","short_pith_number":"pith:FVMKECSP","schema_version":"1.0","canonical_sha256":"2d58a20a4fae175df03d4acd80bbda48bf45972da31a816513c1ca8ae6f201c5","source":{"kind":"arxiv","id":"2306.03341","version":6},"attestation_state":"computed","paper":{"title":"Inference-Time Intervention: Eliciting Truthful Answers from a Language Model","license":"http://creativecommons.org/licenses/by/4.0/","headline":"Shifting activations in a few attention heads during inference raises LLM truthfulness on TruthfulQA from 32.5 percent to 65.1 percent.","cross_cats":["cs.AI","cs.CL"],"primary_cat":"cs.LG","authors_text":"Fernanda Vi\\'egas, Hanspeter Pfister, Kenneth Li, Martin Wattenberg, Oam Patel","submitted_at":"2023-06-06T01:26:53Z","abstract_excerpt":"We introduce Inference-Time Intervention (ITI), a technique designed to enhance the \"truthfulness\" of large language models (LLMs). ITI operates by shifting model activations during inference, following a set of directions across a limited number of attention heads. This intervention significantly improves the performance of LLaMA models on the TruthfulQA benchmark. On an instruction-finetuned LLaMA called Alpaca, ITI improves its truthfulness from 32.5% to 65.1%. We identify a tradeoff between truthfulness and helpfulness and demonstrate how to balance it by tuning the intervention strength. "},"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":true},"canonical_record":{"source":{"id":"2306.03341","kind":"arxiv","version":6},"metadata":{"license":"http://creativecommons.org/licenses/by/4.0/","primary_cat":"cs.LG","submitted_at":"2023-06-06T01:26:53Z","cross_cats_sorted":["cs.AI","cs.CL"],"title_canon_sha256":"807aff8083d26511bae73c8d06f1bdd9b9330faec41de63b95bac2266249f21b","abstract_canon_sha256":"56ab04632bca05f6f1938d141d06fa754445e1e27475c0d20eb27848e8ba8740"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-05-17T23:38:52.913711Z","signature_b64":"NPVF560AReBkQLyrxeoqoHve66t3awk4NGdsq7TneFzQvjj6QsC2osD0FEVyDJ15OGT9Hd/b0xg8Kn+o1pocBQ==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"2d58a20a4fae175df03d4acd80bbda48bf45972da31a816513c1ca8ae6f201c5","last_reissued_at":"2026-05-17T23:38:52.912988Z","signature_status":"signed_v1","first_computed_at":"2026-05-17T23:38:52.912988Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"Inference-Time Intervention: Eliciting Truthful Answers from a Language Model","license":"http://creativecommons.org/licenses/by/4.0/","headline":"Shifting activations in a few attention heads during inference raises LLM truthfulness on TruthfulQA from 32.5 percent to 65.1 percent.","cross_cats":["cs.AI","cs.CL"],"primary_cat":"cs.LG","authors_text":"Fernanda Vi\\'egas, Hanspeter Pfister, Kenneth Li, Martin Wattenberg, Oam Patel","submitted_at":"2023-06-06T01:26:53Z","abstract_excerpt":"We introduce Inference-Time Intervention (ITI), a technique designed to enhance the \"truthfulness\" of large language models (LLMs). ITI operates by shifting model activations during inference, following a set of directions across a limited number of attention heads. This intervention significantly improves the performance of LLaMA models on the TruthfulQA benchmark. On an instruction-finetuned LLaMA called Alpaca, ITI improves its truthfulness from 32.5% to 65.1%. We identify a tradeoff between truthfulness and helpfulness and demonstrate how to balance it by tuning the intervention strength. "},"claims":{"count":4,"items":[{"kind":"strongest_claim","text":"ITI improves truthfulness of an instruction-finetuned LLaMA (Alpaca) on TruthfulQA from 32.5% to 65.1% by shifting activations along learned directions in a limited number of attention heads.","source":"verdict.strongest_claim","status":"machine_extracted","claim_id":"C1","attestation":"unclaimed"},{"kind":"weakest_assumption","text":"That the truthful directions identified from a few hundred examples remain effective and stable across unseen prompts and do not introduce systematic new errors beyond the documented truthfulness-helpfulness tradeoff.","source":"verdict.weakest_assumption","status":"machine_extracted","claim_id":"C2","attestation":"unclaimed"},{"kind":"one_line_summary","text":"ITI shifts activations in limited attention heads using directions found from a few hundred examples, raising Alpaca's TruthfulQA truthfulness from 32.5% to 65.1% while allowing tunable tradeoff with helpfulness.","source":"verdict.one_line_summary","status":"machine_extracted","claim_id":"C3","attestation":"unclaimed"},{"kind":"headline","text":"Shifting activations in a few attention heads during inference raises LLM truthfulness on TruthfulQA from 32.5 percent to 65.1 percent.","source":"verdict.pith_extraction.headline","status":"machine_extracted","claim_id":"C4","attestation":"unclaimed"}],"snapshot_sha256":"63f22d549ebce930542344d33c44ff0d1a92c33dad4d3766ceae0d45f65a21bd"},"source":{"id":"2306.03341","kind":"arxiv","version":6},"verdict":{"id":"e252558a-df02-4a06-a13f-636a94c834b1","model_set":{"reader":"grok-4.3"},"created_at":"2026-05-15T09:16:05.950761Z","strongest_claim":"ITI improves truthfulness of an instruction-finetuned LLaMA (Alpaca) on TruthfulQA from 32.5% to 65.1% by shifting activations along learned directions in a limited number of attention heads.","one_line_summary":"ITI shifts activations in limited attention heads using directions found from a few hundred examples, raising Alpaca's TruthfulQA truthfulness from 32.5% to 65.1% while allowing tunable tradeoff with helpfulness.","pipeline_version":"pith-pipeline@v0.9.0","weakest_assumption":"That the truthful directions identified from a few hundred examples remain effective and stable across unseen prompts and do not introduce systematic new errors beyond the documented truthfulness-helpfulness tradeoff.","pith_extraction_headline":"Shifting activations in a few attention heads during inference raises LLM truthfulness on TruthfulQA from 32.5 percent to 65.1 percent."},"references":{"count":0,"sample":[],"resolved_work":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57","internal_anchors":0},"formal_canon":{"evidence_count":3,"snapshot_sha256":"c68f4ca9d023a5455fd386ede3b40b14ae4ff3075acad25cae7499bf72b1b5e2"},"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":"2306.03341","created_at":"2026-05-17T23:38:52.913073+00:00"},{"alias_kind":"arxiv_version","alias_value":"2306.03341v6","created_at":"2026-05-17T23:38:52.913073+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2306.03341","created_at":"2026-05-17T23:38:52.913073+00:00"},{"alias_kind":"pith_short_12","alias_value":"FVMKECSPVYLV","created_at":"2026-05-18T12:33:33.725879+00:00"},{"alias_kind":"pith_short_16","alias_value":"FVMKECSPVYLV34B5","created_at":"2026-05-18T12:33:33.725879+00:00"},{"alias_kind":"pith_short_8","alias_value":"FVMKECSP","created_at":"2026-05-18T12:33:33.725879+00:00"}],"events":[],"event_summary":{},"paper_claims":[],"inbound_citations":{"count":22,"internal_anchor_count":22,"sample":[{"citing_arxiv_id":"2309.11495","citing_title":"Chain-of-Verification Reduces Hallucination in Large Language Models","ref_index":118,"is_internal_anchor":true},{"citing_arxiv_id":"2309.03883","citing_title":"DoLa: Decoding by Contrasting Layers Improves Factuality in Large Language Models","ref_index":53,"is_internal_anchor":true},{"citing_arxiv_id":"2401.01313","citing_title":"A Comprehensive Survey of Hallucination Mitigation Techniques in Large Language Models","ref_index":45,"is_internal_anchor":true},{"citing_arxiv_id":"2605.12809","citing_title":"Correcting Influence: Unboxing LLM Outputs with Orthogonal Latent Spaces","ref_index":271,"is_internal_anchor":true},{"citing_arxiv_id":"2605.12991","citing_title":"Not Just RLHF: Why Alignment Alone Won't Fix Multi-Agent Sycophancy","ref_index":22,"is_internal_anchor":true},{"citing_arxiv_id":"2311.05232","citing_title":"A Survey on Hallucination in Large Language Models: Principles, Taxonomy, Challenges, and Open Questions","ref_index":179,"is_internal_anchor":true},{"citing_arxiv_id":"2605.11448","citing_title":"Deep Minds and Shallow Probes","ref_index":34,"is_internal_anchor":true},{"citing_arxiv_id":"2605.09773","citing_title":"Exploitation Without Deception: Dark Triad Feature Steering Reveals Separable Antisocial Circuits in Language Models","ref_index":15,"is_internal_anchor":true},{"citing_arxiv_id":"2312.06681","citing_title":"Steering Llama 2 via Contrastive Activation Addition","ref_index":8,"is_internal_anchor":true},{"citing_arxiv_id":"2604.11141","citing_title":"Reducing Hallucination in Enterprise AI Workflows via Hybrid Utility Minimum Bayes Risk (HUMBR)","ref_index":16,"is_internal_anchor":true},{"citing_arxiv_id":"2604.06277","citing_title":"Weakly Supervised Distillation of Hallucination Signals into Transformer Representations","ref_index":10,"is_internal_anchor":true},{"citing_arxiv_id":"2604.13759","citing_title":"The cognitive companion: a lightweight parallel monitoring architecture for detecting and recovering from reasoning degradation in LLM agents","ref_index":6,"is_internal_anchor":true},{"citing_arxiv_id":"2604.15741","citing_title":"Learning Uncertainty from Sequential Internal Dispersion in Large Language Models","ref_index":27,"is_internal_anchor":true},{"citing_arxiv_id":"2604.17663","citing_title":"ATLAS: Constitution-Conditioned Latent Geometry and Redistribution Across Language Models and Neural Perturbation Data","ref_index":14,"is_internal_anchor":true},{"citing_arxiv_id":"2604.19974","citing_title":"Are LLM Uncertainty and Correctness Encoded by the Same Features? A Functional Dissociation via Sparse Autoencoders","ref_index":45,"is_internal_anchor":true},{"citing_arxiv_id":"2604.19018","citing_title":"Local Linearity of LLMs Enables Activation Steering via Model-Based Linear Optimal Control","ref_index":42,"is_internal_anchor":true},{"citing_arxiv_id":"2604.21571","citing_title":"Separable Expert Architecture: Toward Privacy-Preserving LLM Personalization via Composable Adapters and Deletable User Proxies","ref_index":28,"is_internal_anchor":true},{"citing_arxiv_id":"2605.01846","citing_title":"Do Large Language Models Plan Answer Positions? Position Bias in Multiple-Choice Question Generation","ref_index":6,"is_internal_anchor":true},{"citing_arxiv_id":"2605.01167","citing_title":"Minimizing Collateral Damage in Activation Steering","ref_index":3,"is_internal_anchor":true},{"citing_arxiv_id":"2605.03196","citing_title":"Geometric Deviation as an Unsupervised Pre-Generation Reliability Signal: Probing LLM Representations for Answerability","ref_index":8,"is_internal_anchor":true},{"citing_arxiv_id":"2605.03160","citing_title":"Pairwise matrices for sparse autoencoders: single-feature inspection mislabels causal axes","ref_index":10,"is_internal_anchor":true},{"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":40,"is_internal_anchor":true}]},"formal_canon":{"evidence_count":3,"sample":[],"anchors":[]},"links":{"html":"https://pith.science/pith/FVMKECSPVYLV34B5JLGYBO62JC","json":"https://pith.science/pith/FVMKECSPVYLV34B5JLGYBO62JC.json","graph_json":"https://pith.science/api/pith-number/FVMKECSPVYLV34B5JLGYBO62JC/graph.json","events_json":"https://pith.science/api/pith-number/FVMKECSPVYLV34B5JLGYBO62JC/events.json","paper":"https://pith.science/paper/FVMKECSP"},"agent_actions":{"view_html":"https://pith.science/pith/FVMKECSPVYLV34B5JLGYBO62JC","download_json":"https://pith.science/pith/FVMKECSPVYLV34B5JLGYBO62JC.json","view_paper":"https://pith.science/paper/FVMKECSP","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2306.03341&json=true","fetch_graph":"https://pith.science/api/pith-number/FVMKECSPVYLV34B5JLGYBO62JC/graph.json","fetch_events":"https://pith.science/api/pith-number/FVMKECSPVYLV34B5JLGYBO62JC/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/FVMKECSPVYLV34B5JLGYBO62JC/action/timestamp_anchor","attest_storage":"https://pith.science/pith/FVMKECSPVYLV34B5JLGYBO62JC/action/storage_attestation","attest_author":"https://pith.science/pith/FVMKECSPVYLV34B5JLGYBO62JC/action/author_attestation","sign_citation":"https://pith.science/pith/FVMKECSPVYLV34B5JLGYBO62JC/action/citation_signature","submit_replication":"https://pith.science/pith/FVMKECSPVYLV34B5JLGYBO62JC/action/replication_record"}},"created_at":"2026-05-17T23:38:52.913073+00:00","updated_at":"2026-05-17T23:38:52.913073+00:00"}