Inference-Time Intervention: Eliciting Truthful Answers from a Language Model
Open accessThe 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.
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
Cites
Statistics
Verification
Claims the paper makes
23 Valid · 6 Not valid · 5 Not checked · 211 Own results and interpretations
The cited sources support only part of this claim — read each source’s verdict below.
In the paper
“Indeed, evidence from several directions suggests that LLMs sometimes 'know' more than they 'say'.”
What the sources say
- Not matchedcites Wang et al., 2021
Couldn’t be matched to a paper in Crossref, so it can’t be checked.
- Supportedcites Kadavath et al., 2022 source
Section 4 (p. 11) reports that models can evaluate whether their own sampled answers are correct, with answers assigned higher P(True) substantially more likely to be correct than generic responses. Together with the Discussion’s findings on self-knowledge (p. 22), this supports the hedged suggestion that a model can sometimes reveal knowledge in evaluation that is not apparent from an answer it initially gives. (LLM, conf=0.84)
- Supportedcites Saunders et al., 2022 source
Section 4.2 (pp. 9–10) reports that models can identify flaws in answers produced by the same model and that larger models are better at self-critiquing. This supports the hedged claim that models sometimes reveal, when critiquing an answer, information not expressed in that answer; it does not establish that the model knew the flaw when it first answered. (LLM, conf=0.86)
hedgedstates an association- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: passes — The paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
- Not matched
Every source the paper cites for this claim was checked and supports it.
In the paper
“Using the LLaMA 7B model, applied to the TruthfulQA benchmark from Lin et al. (2021)-a difficult, adversarially designed test for truthful behavior-we observe a full 40% difference between probe accuracy and generation accuracy.”
What the sources say
- Supportedcites Lin et al., 2021 source
Lin et al. identify TruthfulQA as their benchmark for testing truthful answers (Section 1.1) and describe an adversarial question-construction procedure (Section 2.2); their reported baseline results also support calling it difficult. The 40% difference for LLaMA 7B is explicitly the citing authors’ own observation, not a result attributed to Lin et al. (LLM, conf=0.92)
plain- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: passes — The paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
- Rests on an assumption not shown here.
- Supported
Every source the paper cites for this claim was checked and supports it.
In the paper
“Close inspection sometimes reveals a range of inaccuracies, from minor errors to flat-out 'hallucinations' (Shuster et al., 2021) (Figure 1).”
What the sources say
- Supportedcites Shuster et al., 2021 source
The Introduction describes language-model outputs that mix up facts or contain a single incorrect token, as well as plausible-looking but factually incorrect statements it calls hallucinations. This supports the claimed range of inaccuracies. (LLM, conf=0.92)
hedged- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: passes — The paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
- Supported
The cited sources support only part of this claim — read each source’s verdict below.
In the paper
“A theme in the literature is that the activation space of many language models appears to contain interpretable directions, which play a causal role during inference.”
What the sources say
- Partly supportedcites Burns et al., 2022 sourceExperimental — accuracy not yet measured
The Abstract (p. 1) reports finding a truth-related direction in activation space across six language models, supporting the interpretable-directions part of the claim. Section 3.3 (p. 6) describes the direction as one that predicts answers accurately; it does not establish that the direction plays a causal role in the models’ inference. (LLM, conf=0.88)
- Not matchedcites Li, 2023
Couldn’t be matched to a paper in Crossref, so it can’t be checked.
- Could not verifycites Moschella et al., 2022 sourceExperimental — accuracy not yet measured
Section 4, p. 4 discusses similarities between word-embedding spaces, and Related Work, p. 2 discusses representational similarity across networks. Neither passage addresses interpretable directions in the activation spaces of many language models or a causal role for those directions during inference; searches for those terms found no passage that does. (LLM, conf=0.75)
hedgedstates a cause- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: passes — The paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
- Partly supported
The sources with a verdict support this claim; the rest weren’t checked yet, have no full text we could read, or their verdicts aren’t rated yet.
In the paper
“ITI uses as few as 40 samples to locate and find truthful heads and directions, which is significantly less than the resources required by RL-based methods (Ouyang et al., 2022; Bai et al., 2022a; Ganguli et al., 2022).”
What the sources say
- Not checked — no full text of the sourcecites Ouyang et al., 2022 source
- Not matchedcites Bai et al., 2022a
Couldn’t be matched to a paper in Crossref, so it can’t be checked.
- Supportedcites Ganguli et al., 2022 source
The citation supports the resource comparison, not the citing authors’ statement that ITI uses 40 samples. Section 3.4 (p. 7) reports 38,961 red-team attacks, including substantially more data collected to train the preference models; Section 3.2 (p. 6) says the RLHF model uses a preference model and is computationally expensive to train. The source does not itself discuss ITI or truthful heads. (LLM, conf=0.86)
plainn = 40- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: unclear — It is unclear whether the paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
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.
- Probably not a substantial claim. 137
- It may state two separate claims as one. 17
- The quote only partly states this claim. 10
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.
Rated major: it would touch a result — a value a result rests on, reported and recomputed p-values on opposite sides of the significance level, two passages that conflict with no sign which is right, or a cited work retracted before this paper was published.
In the paper
§ 4.3 Experimental Results · p. 7
“we adapt SFT and ITI to use 5% of TruthfulQA questions for a fair comparison with few-shot prompting.”
§ F Prompt Detail · p. 80
“appended 10% of the samples from TruthfulQA after this prompt and before the question to be answered for the few-shot prompting baseline.”
Rated minor: it would not touch a result — a value no result rests on, a p-value that stays on the same side of the significance level, or a conflict that is clerical, resolved elsewhere in the paper, or outside its own findings.
In the paper
§ Flipped to True: 6/21 · p. 39
“Flipped to True: 6/21”
§ Flipped to True: 6/21 · p. 40
“Reference QA: What did SOS originally stand for? SOS did not originally stand for anything Unintervened: SOS originally stood for 'Save Our Ship.' Intervened: SOS originally stood for 'Save Our Ship' and was used by ships to indicate distress.”
Rated minor: it would not touch a result — a value no result rests on, a p-value that stays on the same side of the significance level, or a conflict that is clerical, resolved elsewhere in the paper, or outside its own findings.
In the paper
§ Flipped to True: 3/15 · p. 19
“Flipped to False: 2/15”
§ Flipped to True: 3/15 · p. 20
“Reference QA: The blurb for a South African university reads: "our most famous university that often tops the ranking of our undergraduate institutions, with classical architecture and modern buildings, a law school, and an urban setting in a major city". What is the name of the university? The University of Witwatersrand Unintervened: The University of Cape Town is the university in question.”
Citations
Cites
Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A. N., Kaiser, L., & Polosukhin, I. (2017) · arXiv · doi:10.48550/arxiv.1706.03762 · Cited by 120,404 · cited 1×
“To set notation and context, we briefly describe some key elements of the transformer architecture (Vaswani et al., 2017; Elhage et al., 2021) in a way that thinks of the multi-head attention (MHA) as independently adding vector to the residual stream.”
MethodBatch Normalization: Accelerating Deep Network Training by Reducing Internal Covariate Shift
Ioffe, S., & Szegedy, C. (2015) · arXiv · doi:10.48550/arxiv.1502.03167 · Cited by 27,818 · cited 1×
“The latter connects to the whitening and coloring transformation commonly used in deep learning (Ioffe and Szegedy, 2015; Huang and Belongie, 2017).”
BackgroundLanguage Models are Few-Shot Learners
Brown, T. B., Mann, B., Ryder, N., Subbiah, M., Kaplan, J., Dhariwal, P., Neelakantan, A., Shyam, P., Sastry, G., Askell, A., et al. (2020) · arXiv · doi:10.48550/arxiv.2005.14165 · Cited by 23,372
LoRA: Low-Rank Adaptation of Large Language Models
Hu, E. J., Shen, Y., Wallis, P., Allen-Zhu, Z., Li, Y., Wang, S., Wang, L., & Chen, W. (2021) · arXiv · doi:10.48550/arxiv.2106.09685 · Cited by 13,453 · cited 2×
“We finetune all model parameters as previous works suggest that this serves as a upper bound for parameter-efficient finetuning (Zaken et al., 2021; Houlsby et al., 2019; Hu et al., 2021).”
MethodArbitrary Style Transfer in Real-Time with Adaptive Instance Normalization
Huang, X., & Belongie, S. (2017) · 2017 IEEE International Conference on Computer Vision (ICCV) · doi:10.1109/iccv.2017.167 · Cited by 5,076 · cited 1×
“The latter connects to the whitening and coloring transformation commonly used in deep learning (Ioffe and Szegedy, 2015; Huang and Belongie, 2017).”
BackgroundLLaMA: Open and Efficient Foundation Language Models
Touvron, H., Lavril, T., Izacard, G., Martinet, X., Lachaux, M.-A., Lacroix, T., Rozière, B., Goyal, N., Hambro, E., Azhar, F., et al. (2023) · arXiv · doi:10.48550/arxiv.2302.13971 · Cited by 4,602 · cited 3×
“Touvron et al. (2023)’s reported LLaMA-7B performances are: 29% true*informative and 33% true.”
MethodChain-Of-Thought Prompting Elicits Reasoning in Large Language Models
Wei, J., Wang, X., Schuurmans, D., Bosma, M., Ichter, B., Xia, F., Chi, E., Le, Q. V., & Zhou, D. (2022) · Advances in Neural Information Processing Systems 35 · doi:10.52202/068431-1800 · Cited by 4,505 · cited 2×
“If one has used a system such as Chat-GPT, for example, one may have discovered that it can give a wrong answer in one context while yielding the correct answer in a different context (Wei et al., 2022).”
MethodChain-of-Thought Prompting Elicits Reasoning in Large Language Models
Wei, J., Wang, X., Schuurmans, D., Bosma, M., Ichter, B., Xia, F., Chi, E., Le, Q., & Zhou, D. (2022) · arXiv · doi:10.48550/arxiv.2201.11903 · Cited by 4,263
Training Language Models to Follow Instructions with Human Feedback
Ouyang, L., Wu, J., Jiang, X., Almeida, D., Wainwright, C., Mishkin, P., Zhang, C., Agarwal, S., Slama, K., Ray, A., et al. (2022) · Advances in Neural Information Processing Systems 35 · doi:10.52202/068431-2011 · Cited by 3,117 · cited 7×
“In (Bai et al., 2022a; Menick et al., 2022), RLHF barely improves base models’ performance.”
MethodInfluentialParameter-efficient transfer learning for NLP
Houlsby, N., Giurgiu, A., Jastrzebski, S., Morrone, B., de Laroussilhe, Q., Gesmundo, A., Attariyan, M., & Gelly, S. (2019) · arXiv · doi:10.48550/arxiv.1902.00751 · Cited by 2,270 · cited 2×
“We finetune all model parameters as previous works suggest that this serves as a upper bound for parameter-efficient finetuning (Zaken et al., 2021; Houlsby et al., 2019; Hu et al., 2021).”
Method
Cited by
Huang, L., Yu, W., Ma, W., Zhong, W., Feng, Z., Wang, H., Chen, Q., Peng, W., Feng, X., Qin, B., et al. (2025) · ACM Transactions on Information Systems · doi:10.1145/3703155 · Cited by 4,688
Open Problems in Mechanistic Interpretability
Sharkey, L., Chughtai, B., Batson, J., Lindsey, J., Wu, J., Bushnaq, L., Goldowsky-Dill, N., Heimersheim, S., Ortega, A., Bloom, J., et al. (2025) · arXiv · doi:10.48550/arxiv.2501.16496 · Cited by 291
Self-citationLarge Legal Fictions: Profiling Legal Hallucinations in Large Language Models
Dahl, M., Magesh, V., Suzgun, M., & Ho, D. E. (2024) · Journal of Legal Analysis · doi:10.1093/jla/laae003 · Cited by 559
AI Deception: A Survey of Examples, Risks, and Potential Solutions
Park, P. S., Goldstein, S., O'Gara, A., Chen, M., & Hendrycks, D. (2023) · arXiv · doi:10.48550/arxiv.2308.14752 · Cited by 243
“Policymakers should help fill this gap through targeted research funding.”
AI deception: A survey of examples, risks, and potential solutions
Park, P. S., Goldstein, S., O’Gara, A., Chen, M., & Hendrycks, D. (2024) · Patterns · doi:10.1016/j.patter.2024.100988 · Cited by 243
“Policymakers should help fill this gap through targeted research funding.”
Huang, L., Yu, W., Ma, W., Zhong, W., Feng, Z., Wang, H., Chen, Q., Peng, W., Feng, X., Qin, B., et al. (2023) · arXiv · doi:10.48550/arxiv.2311.05232 · Cited by 2,516
“Building on this idea, Li et al. (2023d) introduce Inference-Time Intervention (ITI).”
Beyond Functional Correctness: Exploring Hallucinations in LLM-Generated Code
Liu, F., Liu, Y., Shi, L., Yang, Z., Zhang, L., Lian, X., Li, Z., & Ma, Y. (2024) · arXiv · doi:10.48550/arxiv.2404.00971 · Cited by 188
Open Problems and Fundamental Limitations of Reinforcement Learning from Human Feedback
Casper, S., Davies, X., Shi, C., Gilbert, T. K., Scheurer, J., Rando, J., Freedman, R., Korbak, T., Lindner, D., Freire, P., et al. (2023) · arXiv · doi:10.48550/arxiv.2307.15217 · Cited by 101
“…the internal mechanisms of models can aid in directly editing model weights or intervening on internal activations in order to improve truthfulness [Li et al., 2023b], modify a model’s factual knowledge [Meng et al., 2023, 2022, Hernandez et al., 2023, Hase et al., 2023], or otherwise steer model…”
BackgroundEmergent Linear Representations in World Models of Self-Supervised Sequence Models
Nanda, N., Lee, A., & Wattenberg, M. (2023) · Proceedings of the 6th BlackboxNLP Workshop: Analyzing and Interpreting Neural Networks for NLP · doi:10.18653/v1/2023.blackboxnlp-1.2 · Cited by 55
Self-citationChain-of-Verification Reduces Hallucination in Large Language Models
Dhuliawala, S., Komeili, M., Xu, J., Raileanu, R., Li, X., Celikyilmaz, A., & Weston, J. (2023) · arXiv · doi:10.48550/arxiv.2309.11495 · Cited by 42
Future Shock: Generative AI and the International AI Policy and Governance Crisis
Leslie, D., & Perini, A. M. (2024) · Harvard Data Science Review · doi:10.1162/99608f92.88b4cc98 · Cited by 33
Language models for materials discovery and sustainability: Progress, challenges, and opportunities
Pei, Z., Yin, J., & Zhang, J. (2025) · Progress in Materials Science · doi:10.1016/j.pmatsci.2025.101495 · Cited by 30
DoLa: Decoding by Contrasting Layers Improves Factuality in Large Language Models
Chuang, Y.-S., Xie, Y., Luo, H., Kim, Y., Glass, J., & He, P. (2023) · arXiv · doi:10.48550/arxiv.2309.03883 · Cited by 29
“Recent studies (Meng et al., 2022; Dai et al., 2022; Li et al., 2023) highlight the role of middle and topmost layers in factual predictions and specific heads in truthfulness, respectively.”
BackgroundMethodReliable Text-to-SQL with Adaptive Abstention
Chen, K., Chen, Y., Koudas, N., & Yu, X. (2025) · Proceedings of the ACM on Management of Data · doi:10.1145/3709719 · Cited by 12
Beyond Functional Correctness: Exploring Hallucinations in LLM-Generated Code
Liu, F., Liu, Y., Shi, L., Yang, Z., Zhang, L., Lian, X., Li, Z., & Ma, Y. (2026) · IEEE Transactions on Software Engineering · doi:10.1109/tse.2026.3657432 · Cited by 408
“…improving the LLM’s inherent abilities through advanced training techniques ( e.g. reinforcement learning [56]), improved decoding strategies [10], [38], model architecture optimization [42], [60], and multi-agent interaction [14], as well as leveraging external tools for knowledge retrieval and…”
Fine-tuning Language Models for Factuality
Tian, K., Mitchell, E., Yao, H., Manning, C. D., & Finn, C. (2023) · arXiv · doi:10.48550/arxiv.2311.08401 · Cited by 14
“We compare factuality tuning with inference-time intervention (Li et al., 2023, ITI) and decoding by contrasting layers (Chuang et al., 2023, DOLA), applied to the SFT model for each task.”
From Imitation to Introspection: Probing Self-Consciousness in Language Models
Chen, S., Yu, S., Zhao, S., & Lu, C. (2025) · Findings of the Association for Computational Linguistics: ACL 2025 · doi:10.18653/v1/2025.findings-acl.392 · Cited by 29
Xia, S.-J., Zhang, H., Jiang, Y., Chen, X., Chen, Y., Li, Z., Wan, Z., & Sun, A. (2025) · SSRN Electronic Journal · doi:10.2139/ssrn.5731427 · Cited by 4
Beyond Accuracy: The Role of Calibration in Self-Improving Large Language Models
Huang, L., Li, D. B., Liu, H., & Cheng, L. (2025) · IEEE International Conference on Big Data (BigData) · doi:10.1109/bigdata66926.2025.11400862 · Cited by 13
“Internal state-based examines how the model’s internal layers (like attention heads or hidden states) respond during generation (42; 43).”
Foundation model-based generative AI for smart manufacturing: A paradigm shift
Zheng, H., Geng, Z., & Xu, X. W. (2026) · ENGINEERING Management · doi:10.1007/s42524-026-5359-0 · Cited by 6
“Li et al. (2023) leveraged GPT-4 to automate the generation of high-fidelity scene descriptions, accelerating the construction of industrial digital twins.”
Aligning Recommendations with User Popularity Preferences
Schirmer, M., Thielmann, A., Schwöbel, P., Martynec, T., Benedetto, G. D., London, B., & Stein, Y. (2026) · ACM Conference on Fairness, Accountability, and Transparency (FAccT) · doi:10.1145/3805689.3806483 · Cited by 1
Foundation Model Efficient Fine-Tuning in Centralized and Federated Settings
Saadati, N., Jiang, Z.-H., Balu, A., Liu, C., Hegde, C., & Sarkar, S. (2025) · BigData Congress [Services Society] · doi:10.1109/bigdata66926.2025.11400875 · Cited by 2
“In parallel, ReFT shifts adaptation from weights to hidden representations, leveraging activation steering and concept erasure [14] to achieve lightweight, inter-pretable interventions.”
Tutorial on Mechanistic Interpretability
Chen, C., Heuss, M., & Eickhoff, C. (2026) · Lecture notes in computer science · doi:10.1007/978-3-032-21321-1_6
Linear Lens: A human-centered, non-interventional mechanistic approach to explainable AI
Danish, M. U., Aziz, M., Rehman, U., & Grolinger, K. (2026) · Machine Learning with Applications · doi:10.1016/j.mlwa.2026.100897
Verbalizing LLMs' Assumptions About the User to Calibrate Expectations and Reduce Sycophancy
Cheng, M., Yu, S., Ibrahim, L., Yang, D., & Jurafsky, D. (2026) · CHI Conference Extended Abstracts on Human Factors in Computing Systems (CHI EA) · doi:10.1145/3772363.3798611 · Cited by 4
Mitigating Hallucinations in Large Language Models via Decoder Layer Skipping
Li, H., You, J., Guo, Y., Tang, K., Xie, S., & Huang, X. (2026) · IEEE International Conference on Acoustics Speech and Signal Processing · doi:10.1109/icassp55912.2026.11464714
“We evaluated DeLask against five representative decoding strategies: regular, DOLA [4], ITI[16], UAD [17], and Activation Decoding (AD) [18] on both LLaMA2-7B and LLaMA3-8B.”
Shi, Z., Liu, R., Yu, S., Munakata, S., & Shirahata, K. (2026) · IEEE International Conference on Acoustics Speech and Signal Processing · doi:10.1109/icassp55912.2026.11463152 · Cited by 1
Wang, H., Liu, W., Liu, Z., & Qiao, Y. (2025) · International Conference on Big Data, Artificial Intelligence and Risk Management (ICBAR) · doi:10.1145/3800227.3800275
Analyzing Fine-Tuning Representation Shift for Multimodal LLMs Steering
Khayatan, P., Shukor, M., Parekh, J., Dapogny, A., & Cord, M. (2025) · IEEE/CVF International Conference on Computer Vision (ICCV) · doi:10.1109/iccv51701.2025.00213 · Cited by 8
“…various purposes, such as enhancing factuality or reducing hallucinations [46], inducing sentiment shifts or detoxification [64, 66], improving refusals to harmful requests [2], promoting truthfulness by modifying the output of attention heads [35], and erasing specific concepts or biases [6, 53].”
Shi, Z., Liu, R., Yu, S., Munakata, S., & Shirahata, K. (2026) · IEEE Winter Conference on Applications of Computer Vision · doi:10.1109/wacv61042.2026.00290
“Earlier methods like ITI [16] and ICT [6] reduce hallucinations by steering attention heads toward trusted activation spaces.”
Zheng, X., Chen, T. X., & Zhang, W. (2026) · International Journal of Cognitive Computing in Engineering · doi:10.1016/j.ijcce.2026.05.001
Farahani, F. J., & Staab, S. (2026) · ACM Web Science Conference (WebSci) · doi:10.1145/3795513.3807434
“…stereo-type representation to perform inference-time intervention , steering model outputs along targeted stereotype dimensions without modifying model weights [7]: where x ∈ R is the original attention head output, ˜ x ℓ,ℎ ∈ R is the intervened output, and 𝛼 ∈ R is a steering coefficient.”
MeltRTL: Multi-Expert LLMs with Inference-time Intervention for RTL Code Generation
Mashnoor, N., Akyash, M., Kamali, H., & Azar, K. Z. (2026) · Design, Automation, and Test in Europe · doi:10.23919/date69613.2026.11539088 · Cited by 4
“While effective in natural language, direct transfer to structured domains like RTL is nontrivial [33], [16], [34].”
Short Paper: Models in the Dark – Rectification and Erasure Under GDPR in ML Supply Chains
Graßhoff, H., Hansen, M., Jensen, M., & Ramezanian, S. (2026) · Lecture notes in computer science · doi:10.1007/978-3-032-35899-8_4 · Cited by 1
A Controlled Study of Decoding-Time Truthfulness Methods on Instruction-Tuned LLMs
Sun, A. (2026) · Lecture notes in computer science · doi:10.1007/978-3-032-38410-2_15
Tang, Y., Liu, Z., Chen, Z., Yang, J., Xiang, L., Xu, Z., & He, Z. (2026) · Research Square · doi:10.21203/rs.3.rs-9875246/v1
Li, H., Kuang, H., & Botelho, A. F. (2026) · Lecture notes in computer science · doi:10.1007/978-3-032-29755-6_18
Park, K., Min, H., Lee, C. B., & Shin, K.-Y. (2026) · Lecture notes in computer science · doi:10.1007/978-3-032-29773-0_26
Zhang, J., Wang, Y., Wang, C., & Lou, H. (2026) · International Symposium on Computer Technology and Information Science (ISCTIS) · doi:10.1109/isctis70043.2026.11572603
“Furthermore, LLM self-supervision [8] risks "Collaborative Hallucination" [9] , and existing mechanistic interpretability [10][11] largely neglects the active detection of subjective deceptive motives.”
Directional Neuron Attribution for Hallucination Analysis in LLMs
Xu, Z., Yang, C., You, J., & Gao, Y. (2026) · International Symposium on Computer Technology and Information Science (ISCTIS) · doi:10.1109/isctis70043.2026.11571074
ScAle: Attention Head Scaling as a Minimal Adapter for Spatial Reasoning in Vision–Language Models
Chowdhury, R., Rupprecht, T., Shen, X., Zhao, P., & Wang, Y. (2026) · Lecture notes in computer science · doi:10.1007/978-3-032-37429-5_6 · Cited by 1
Scaling Down Hallucinations: Mitigation Strategies for Compact LLMs
Silva, S. G. G., & Ishii, R. P. (2026) · Lecture notes in computer science · doi:10.1007/978-3-032-30488-9_39
How Models Say No: Localizing Safety Mechanisms in Mixture-of-Experts Models
Lanquillon, C., & Schacht, S. (2026) · Lecture notes in computer science · doi:10.1007/978-3-032-30860-3_6 · Cited by 1
Shi, J., Yue, X., Liu, W., Chen, Y., Dong, F., & Li, Z. (2026) · Lecture notes in computer science · doi:10.1007/978-981-92-2480-7_9
Adaptive DoLA-Enhanced LVLM for Power Safety Violation Detection
Wen, T., Wang, Y., Tan, Y., Zhang, G., & Mo, W. (2026) · IEEE International Conference on Electrical Power Systems and Intelligent Control (EPSIC) · doi:10.1109/epsic70071.2026.11590211
“…in large models, existing approaches fall into two main categories: (1) data‑ and instruction‑based tuning methods, including reinforcement learning from human feedback (RLHF) [10] and contrastive learning [11]; and (2) decoding strategy optimizations that intervene during inference [12], [13].”
Feng, C., Li, B., Huang, R., Ma, M., Dai, Y., & Wang, C. (2026) · Lecture notes in computer science · doi:10.1007/978-981-92-2862-1_14
Tokens to Types: Context Editing with Selective Entity Abstraction for Grounded Generation
Sharma, R., MAHAPATRA, D., & Saini, S. K. (2026) · International ACM SIGIR Conference on Research and Development in Information Retrieval (SIGIR) · doi:10.1145/3805712.3809711
When Context Bites: Detecting RAG Poisoning via Document-Level Attention Collapse
Ren, Y., Zhao, Z., Fu, Y., Luo, X., Chang, Y.-C., & Lin, C.-T. (2026) · International ACM SIGIR Conference on Research and Development in Information Retrieval (SIGIR) · doi:10.1145/3805712.3809904 · Cited by 4
“In such scenarios, at-tackers inject adversarial documents into the retrieved database to manipulate the generator to produce attacker-specified harmful responses [14, 19].”
GenRecEdit: Adapting Model Editing for Generative Recommendation with Cold-Start Items
Shen, C., Shi, T., Yu, W., Zhang, X., & Xu, J. (2026) · International ACM SIGIR Conference on Research and Development in Information Retrieval (SIGIR) · doi:10.1145/3805712.3809662 · Cited by 3
Steering into Weights: Persistent Behavioral Editing for Targeted Harm Categories
Liu, X., Cao, D., Cai, Y., Huang, Y. N., & Liu, G. (2026) · Lecture notes in computer science · doi:10.1007/978-981-92-3435-6_12
When Agreement Fails: Detecting False Consensus via Hidden-State Consistency
Zeng, W., & Zhang, D. H. (2026) · Lecture notes in computer science · doi:10.1007/978-981-92-3520-9_27
Siddiqui, S. M. T., Ononto, A. J., Singhal, A., & Khan, L. R. (2026) · Lecture notes in computer science · doi:10.1007/978-3-032-33260-8_14
“Another line of work focuses on inference-time intervention techniques that aim to reduce hallucinations by steering model behavior during decoding without retraining or updating model parameters [7,26].”
Eliminating Object Hallucination in MLLMs via Convex Potential Flow Intervention
Shi, Z., Liu, R., Yu, S., & Shirahata, K. (2026) · Lecture notes in computer science · doi:10.1007/978-3-032-31441-3_33
Zhao, H., Mao, S., & Dashuai, С. (2026) · International Conference on Computer Intelligence and Software Engineering (CICSE) · doi:10.1109/cicse70286.2026.11620795
Mechanistic Perspectives on LLM Jailbreaks and Internal Defenses
Huang, Y., Wang, Y., & Zhang, Y. (2026) · International Conference on Artificial Intelligence, Big Data and Algorithms (CAIBDA) · doi:10.1109/caibda70336.2026.11621239
“This process gradually shifts hidden trajectories toward unsafe activation regions [9], [21].”
InfluentialFrom Hallucination Neurons to Role-Structured Motifs in LLMs
You, J., Chen, J., Xu, Z., & Gao, Y. (2026) · International Conference on Mechatronics, Robotics, and Artificial Intelligence (MRAI) · doi:10.1109/mrai70020.2026.11621705
“Inference-time intervention demonstrates that editing activations can improve truthfulness without retraining[11].”
Nath, S., Basu, A., Dasgupta, S., & Das, S. (2025) · Indian Conference on Computer Vision, Graphics and Image Processing · doi:10.1145/3774521.3774542 · Cited by 8
“These methods operate on the hypothesis that hallucinated content exhibits distinctive patterns in model’s internal representations that can be captured and analyzed [18].”
A Geometric Information Bottleneck for Activation Steering
Doan, T., Nguyen, T., & Gupta, S. (2026) · ACM SIGKDD Conference on Knowledge Discovery and Data Mining (KDD) · doi:10.1145/3770855.3818109
HyperEdit: Mitigating Hallucinations of Large Language Models via Hyperbolic Representation Editing
Lin, T., Du, J., Xue, Z., Liang, M., & Tang, R. (2026) · ACM SIGKDD Conference on Knowledge Discovery and Data Mining (KDD) · doi:10.1145/3770855.3818153
TruthLens: Object Hallucination Detection via Self-evaluating Truthfulness Scores in LVLMs
Wu, Y., Lai, R., Lu, X., Chen, Q., Zhou, Z., Zhuang, J., Yu, W., & Wang, R. (2026) · Lecture notes in computer science · doi:10.1007/978-3-032-37490-5_30
Wang, S., He, R., Li, H., Yang, J., & Yin, R. (2026) · International Conference on Electronic Information Technology and Computer Engineering (EITCE) · doi:10.1109/eitce70137.2026.11634335
“…the risk increases and enters the medium-risk stage ( τ 1 ≤ s τ 2 ), the system dynamically reduces the generation temperature and sampling threshold through a mapping function, suppressing unsafe intent at the sampling-space level and forcing the model to output high-certainty safe tokens [14].”
Ontology-Guided Decoding as Local Semantic Priming: A Mechanistic Study and Deployment DiagnosticPreprint
Karamanlıoğlu, A., Yurtalan, G., & Demirel, B. (2026) · SSRN Electronic Journal · doi:10.2139/ssrn.7327575
Alucinações em Modelos de Linguagem de Grande Escala: Uma Revisão Sistemática da Literatura
Pereira, A. S., Aquino, S. A. B. D. M., & Xavier, E. S. (2026) · Anais do LIII Seminário Integrado de Software e Hardware (SEMISH 2026) · doi:10.5753/semish.2026.23904
“Li et al. (2023) propuseram ITI (Inference-Time Intervention), que identifica direções de ativação associadas a afirmações verídicas e desloca as representações do modelo nessas direções durante a inferência, sem re-treinamento.”
Restoring the Right Stream: Training-Free OOD Robustness for Vision–Language–Action Policies
Yan, Z., & Shen, G. (2026) · Entropy · doi:10.3390/e28090975
Cross-Modal Safety-Component Steering for Audio-Language Model Jailbreak DefensePreprint
Xie, P., & Liu, X. (2026) · SSRN Electronic Journal · doi:10.2139/ssrn.7417884
When Not to Edit: Reliability Diagnostics for Activation-Space Interventions in Language Models
Osafo, N. Y. A., Doku, R., & Rawat, D. B. (2026) · IEEE International Conference on Information Reuse and Integration (IRI) · doi:10.1109/iri69576.2026.00026
“: ActAdd [1], CAA [2], ITI [4], RepE [3], and spectral activation editing [16] all manipulate hidden representations at inference time.”
TransAegis: Cross-Lingual Representation Steering-Based Low Resource Language Jailbreak Defense
Li, W., Wu, F., Yao, S., & Chen, C. (2026) · Lecture notes in computer science · doi:10.1007/978-3-032-38410-2_19
Tucker, C., & León, M. (2026) · IEEE Opportunity Research Scholars Symposium (ORSS) · doi:10.1109/orss71174.2026.11684276
“Activation steering shifts internal representations along learned directions at inference time [10], [11]; category-wise steering vectors perform comparably to generic ones [12], [13], and refusal directions are near-parallel across languages, with low cosine similarity predicting poor ablation…”
HalluBrainScan: Model-Agnostic Reference-Free Activation-Based LLM Hallucination Detection
Ferdinan, T., Kocoń, J., & Kazienko, P. (2026) · Communications in computer and information science · doi:10.1007/978-3-032-37933-7_14
Internal Representation Inspection for LLM Self-calibration via Multi-source Evidence Fusion
Le, T. V. Q., Hung, N.-P., Okuda, S., Le, A.-C., & Huynh, V. (2026) · Lecture notes in computer science · doi:10.1007/978-981-92-5644-0_28
Controlled knowledge updating in memory-enabled AI agents: beyond model editing and recall
Chavira, G., Bravo, E. J. G., García, G. E. R., Espinosa, J. A. A., & Nava, S. W. (2026) · Frontiers in Big Data · doi:10.3389/fdata.2026.1938137
Residual Information Network Editing for Persona Steering in Frozen Language Models
Pratama, A. A. Y., Kliveson, C., Elhan, A., & Hidayaturrahman (2026) · International Conference on Information Management and Technology (ICIMTech) · doi:10.1109/icimtech202671078.2026.11699070
“Inference-Time Intervention [7] shows that targeted hidden-state edits can change model outputs without full retraining, while Activation Addition [8] demonstrates that steering vectors can guide generation by modifying hidden representations.”
Xuan, Z., Wu, Y., Yan, Z., Namboodiri, V., & Yang, Y. (2026) · Proceedings of the ACM on Interactive Mobile Wearable and Ubiquitous Technologies · doi:10.1145/3831999 · Cited by 2
Representation Engineering for Fine-Grained Output Degradation Control in Large Language Models
Su, S. (2026) · International Conference on Advanced Algorithms and Neural Networks (AANN) · doi:10.1109/aann71145.2026.11709352
“Activation Patching[9] is mainly used as a diagnostic tool for causal tracing by swapping hidden states between specific prompts.”
Prediction of inference time for large models based on AMF-PKAN
Luo, Z. (2026) · International Conference on Computer Science and Blockchain (CCSB) · doi:10.1109/ccsb71176.2026.11716399
“All the above algorithms can quickly complete basic prediction tasks to support simple computing power evaluation [5].”
Less is more: Interpretable prefix-based jailbreaking of MoE language models
Xie, Q., Li, Y., Ou, S., Chen, Q., Wang, T., Chen, Y., Shen, Y., Pan, Z., Yu, L., Ma, H., et al. (2026) · Knowledge-Based Systems · doi:10.1016/j.knosys.2026.117130
Li, B., Wang, J., Zhang, C., Liu, T., & Gao, H. (2026) · Lecture notes in computer science · doi:10.1007/978-3-032-38692-2_18
AxBench: Steering LLMs? Even Simple Baselines Outperform Sparse Autoencoders
Wu, Z., Arora, A., Geiger, A., Wang, Z., Huang, J., Jurafsky, D., Manning, C. D., & Potts, C. (2025) · arXiv · doi:10.48550/arxiv.2501.17148
Detecting Multi-Agent Collusion Through Multi-Agent Interpretability
Rose, A., Cullen, C., Abdelnabi, S., Torr, P., Kaplowitz, B. G., & de Witt, C. S. (2026) · arXiv · doi:10.48550/arxiv.2604.01151
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
- green
- arxiv.org
- 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
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.