Student or educator? Sign up with your school email and get 6 months of Pro free.

The DOI matches its Crossref record; 57 of the 161 references it cites were found in a registry or an index; 33 of the 49 sources its claims cite were checked for support.

Inference-Time Intervention: Eliciting Truthful Answers from a Language Model

Open access
Li, Kenneth, Patel, Oam, Viégas, Fernanda, Pfister, Hanspeter, Wattenberg, Martin·2023·Advances in Neural Information Processing Systems 36

Abstract

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. ITI is minimally invasive and computationally inexpensive. Moreover, the technique is data efficient: while approaches like RLHF require extensive annotations, ITI locates truthful directions using only few hundred examples. Our findings suggest that LLMs may have an internal representation of the likelihood of something being true, even as they produce falsehoods on the surface.

Overview

CoChat Summary

TL;DR

Inference-Time Intervention (ITI) shifts model activations during inference along truth-correlated directions in a sparse set of attention heads, improving truthfulness on TruthfulQA. On Alpaca, ITI raises true*informative from 32.5% to 65.1%, while revealing a trade-off between truthfulness and helpfulness.

Key findings

  • ITI substantially improves TruthfulQA performance: Alpaca true*informative rises from 32.5% to 65.1% and true from 32.7% to 66.6%; Vicuna true*informative rises from 51.5% to 74.0% and true from 55.6% to 88.6%.
  • On LLaMA-7B, baseline true*informative is 30.5% and true is 31.6%; Baseline + ITI reaches 43.5% true*informative and 49.1% true, while Few-shot Prompting + ITI reaches 51.4% true*informative and 53.5% true.
  • The paper reports a 40% difference between probe accuracy and generation accuracy on TruthfulQA with LLaMA-7B, and the top probing head (18th head in 14th layer) has 83.3% validation accuracy.
  • Out-of-distribution generalization is smaller but positive: Natural Questions 46.6 to 51.3, TriviaQA 89.6 to 91.1, and MMLU 35.71 to 40.16.

Methods

  • ITI first trains linear probes on attention-head activations from TruthfulQA to rank heads by truthfulness-related validation accuracy, then during inference shifts activations along selected truthful directions for the top K heads, repeated autoregressively.
  • Experiments use LLaMA, Alpaca, and Vicuna; TruthfulQA contains 817 questions reorganized into N = 5,918 QA pairs with binary truthfulness labels. The main hyperparameters are K and α, with optimal values reported as K = 48 and α = 15.
  • Baselines include supervised fine-tuning, few-shot prompting, and instruction fine-tuning; intervention directions compared include probe weight direction, mass mean shift, CCS, and random directions. Metrics include true*informative, true, multiple-choice accuracy, cross entropy, and KL divergence.

Limitations

  • The paper states it does not claim ITI by itself is anywhere near sufficient for ensuring truthful answers from LLMs, and it makes no claims about mechanistically understanding what ITI does to internal representations.
  • The paper states TruthfulQA does not cover the full range of meanings of “truth,” and it identifies a trade-off between truthfulness and helpfulness that must be tuned by the user.
  • Evident from the text: evaluation relies heavily on GPT-judge, though the authors report human evaluation found truthfulness slightly overestimated and informativeness underestimated; generalization tests are limited to a few benchmarks and 7B-scale models, with hyperparameters and directions learned from TruthfulQA.

Model-written summary of a shortened version of the paper’s full text (it is longer than we read). Check it against the paper before relying on it.

Integrity signals

Registry

Crossref · 10.52202/075280-1797

Cites

57 of 57 references verified0 citation errors · 0 not checkable

Statistics

Nothing to recompute.

Verification

Claims the paper makes

23 Valid · 6 Not valid · 5 Not checked · 211 Own results and interpretations

91 more claims · 1 with their cited sources checked

Create a free account to see all 245 claims and to verify the 1 reference still open.

See all claims — free account
  • Probably not a substantial claim. 137
  • It may state two separate claims as one. 17
  • The quote only partly states this claim. 10

Create a free account to see these claims.

Self-check

Percentages, sums and reported differences re-computed from the paper’s own numbers.

Each contradiction between two passages listed here was confirmed by an independent check.

3 of 3

Citations

Cites

Show all 161 references

Cited by

See all 117

Metadata

Authors
Li, K., Patel, O., Viégas, F., Pfister, H., & Wattenberg, M.
Venue
Advances in Neural Information Processing Systems 36
Published
2023
Pages
41451-41530
ISSN
1049-5258 (ISSN-L)
OpenAlex
W7133235880
Cited by
117
References
161
Citations by year
2026: 64 · 2025: 29 · 2024: 20 · 2023: 4
FWCI
58.47
Type
conference paper
Language
en
Open access
Licence
cc-by
Locations
  • neural information processing systems (publishedVersion)
  • arXiv (Cornell University) (submittedVersion, cc-by)
  • arXiv (Cornell University) (cc-by)
Funders
National Science Foundation (1901030, 2124179)
Topics
  • Social Sciences › Psychology › Social Psychology › Deception detection and forensic psychology
  • Physical Sciences › Computer Science › Artificial Intelligence › Topic Modeling
  • Social Sciences › Sociology and Political Science › Misinformation and Its Impacts
Keywords
inference-time intervention, language models, hallucination mitigation

About these checks

Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.

The claim verdicts this report counts are experimental: their accuracy is not yet measured.

A language model drew the claims from the paper on 2026-10-27, under version 3 of the claim definition (model deepseek-v4.1-flash, prompt extract-v3).

A second model labelled each claim on 2026-10-27 (model jev-1.13-20260917, definition revision gold-guideline-v0.2).

Cited sources are checked with a language model’s help (gpt-6-sol).

Statistics and arithmetic are recomputed, not judged.

Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.

What counts as a claim ›How we measure confidence ›How we check this ›

Are you an author, or do you see a problem with this page? Request a review

How to cite

Cite CoChat for the checks on this page, and the paper for its own findings.

This page
CoChat. (2026). Citation check: Inference-Time Intervention: Eliciting Truthful Answers from a Language Model. https://app.cochat.ai/papers/10.52202/075280-1797
The paper
Li, K., Patel, O., Viégas, F., Pfister, H., & Wattenberg, M. (2023). Inference-Time Intervention: Eliciting Truthful Answers from a Language Model. Advances in Neural Information Processing Systems 36. https://doi.org/10.52202/075280-1797

Data sources

Bibliographic data: Semantic Scholar, OpenAlex, Crossref.

Contains information from the Semantic Scholar Academic Graph Datasets, made available under the ODC Attribution License (ODC-BY 1.0).

Semantic Scholar’s data is described in: Kinney, R., et al. (2023). The Semantic Scholar Open Data Platform. arXiv:2301.10140.