{"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"}