CoChat

Survey and analysis of hallucinations in large language models: attribution to prompting strategies or model behavior

Open access
Anh-Hoang, Dang, Tran, Vu, Nguyen, Le-Minh · 2025 · Frontiers in Artificial Intelligence

The DOI matches its Crossref record; 39 of the 40 references it cites were found in a registry or an index.

Abstract

Hallucination in Large Language Models (LLMs) refers to outputs that appear fluent and coherent but are factually incorrect, logically inconsistent, or entirely fabricated. As LLMs are increasingly deployed in education, healthcare, law, and scientific research, understanding and mitigating hallucinations has become critical. In this work, we present a comprehensive survey and empirical analysis of hallucination attribution in LLMs. Introducing a novel framework to determine whether a given hallucination stems from not optimize prompting or the model's intrinsic behavior. We evaluate state-of-the-art LLMs—including GPT-4, LLaMA 2, DeepSeek, and others—under various controlled prompting conditions, using established benchmarks (TruthfulQA, HallucinationEval) to judge factuality. Our attribution framework defines metrics for Prompt Sensitivity (PS) and Model Variability (MV) , which together quantify the contribution of prompts vs. model-internal factors to hallucinations. Through extensive experiments and comparative analyses, we identify distinct patterns in hallucination occurrence, severity, and mitigation across models. Notably, structured prompt strategies such as chain-of-thought (CoT) prompting significantly reduce hallucinations in prompt-sensitive scenarios, though intrinsic model limitations persist in some cases. These findings contribute to a deeper understanding of LLM reliability and provide insights for prompt engineers, model developers, and AI practitioners. We further propose best practices and future directions to reduce hallucinations in both prompt design and model development pipelines.

Overview

The claims and statistics of this paper haven’t been checked yet.

Free: it reads the open-access text and checks the first 5 claims that cite a source we can check. Create a free account to check every claim.

A few free checks a day · papers already verified open instantly · results are public

Integrity signals

Registry

Crossref · 10.3389/frai.2025.1622292

Cites

39 of 40 references verified0 citation errors · 1 not checkable

Statistics

Figures

Figure 1 Attribution quadrants defined using median cutoffs for Prompt Sensitivity (PS = 0.079) and Model Variability (MV = 0.057), computed across all evaluated models in Table 4. The quadrants correspond to prompt-dominant (high PS, low MV), model-dominant (low PS, high MV), mixed-origin (high PS and MV), and unclassified (low PS and MV) hallucinations. Using medians provides an objective, distribution-aware threshold rather than arbitrary cutoffs.
Figure 1 Attribution quadrants defined using median cutoffs for Prompt Sensitivity (PS = 0.079) and Model Variability (MV = 0.057), computed across all evaluated models in Table 4. The quadrants correspond to prompt-dominant (high PS, low MV), model-dominant (low PS, high MV), mixed-origin (high PS and MV), and unclassified (low PS and MV) hallucinations. Using medians provides an objective, distribution-aware threshold rather than arbitrary cutoffs.CC BY · PubMed Central
Figure 2 Free experimental pipeline: dataset → prompting → model generation → evaluation → attribution analysis.
Figure 2 Free experimental pipeline: dataset → prompting → model generation → evaluation → attribution analysis.CC BY · PubMed Central
Figure 3 Mean ± SD across 3 seeds × 5 prompt variants; n = 100 examples/model. CoT reduces hallucinations most consistently.
Figure 3 Mean ± SD across 3 seeds × 5 prompt variants; n = 100 examples/model. CoT reduces hallucinations most consistently.CC BY · PubMed Central
Figure 4 Radar plot using polaraxis. Axes show Factuality, Coherence, Prompt Sensitivity, Model Variability, and Usability. Radial scale is normalized (0–1; higher is better). Semi-transparent fills reveal overlaps; dotted crosshairs aid reading.
Figure 4 Radar plot using polaraxis. Axes show Factuality, Coherence, Prompt Sensitivity, Model Variability, and Usability. Radial scale is normalized (0–1; higher is better). Semi-transparent fills reveal overlaps; dotted crosshairs aid reading.CC BY · PubMed Central
Figure 5 Distributions of Prompt Sensitivity (PS) and Model Variability (MV). Vertical dashed lines indicate median cutoffs (PS = 0.079, MV = 0.057), which are used to define “low” vs. “high” thresholds in the attribution quadrants (Figure 1). This ensures that quadrant categorization is aligned with the actual distributions of PS and MV across evaluated models.
Figure 5 Distributions of Prompt Sensitivity (PS) and Model Variability (MV). Vertical dashed lines indicate median cutoffs (PS = 0.079, MV = 0.057), which are used to define “low” vs. “high” thresholds in the attribution quadrants (Figure 1). This ensures that quadrant categorization is aligned with the actual distributions of PS and MV across evaluated models.CC BY · PubMed Central
Figure 6 Attribution quadrants based on Prompt Sensitivity (PS) and Model Variability (MV). Vertical and horizontal dashed lines represent median cutoffs (PS = 0.079, MV = 0.057). Models are positioned by their measured PS and MV scores. Quadrant boundaries define attribution categories (Prompt-dominant, Model-dominant, Mixed-origin, Unclassified), consistent with Table 4.
Figure 6 Attribution quadrants based on Prompt Sensitivity (PS) and Model Variability (MV). Vertical and horizontal dashed lines represent median cutoffs (PS = 0.079, MV = 0.057). Models are positioned by their measured PS and MV scores. Quadrant boundaries define attribution categories (Prompt-dominant, Model-dominant, Mixed-origin, Unclassified), consistent with Table 4.CC BY · PubMed Central
Figure 7 Prompt sensitivity–model variability facets with shared HR encoding. Each panel corresponds to a prompting technique (Zero-shot, Few-shot, CoT, Instruction, Vague). Model locations are fixed across panels by Prompt Sensitivity (PS, x-axis) and Model Variability (MV, y-axis); dashed lines mark the median cutoffs (PS = 0.079, MV = 0.057). Hallucination Rate (HR, %) is encoded by marker size (size ∝ HR) and a shared colorbar (identical range in all panels). Values are Mean ± SD over three seeds × five prompt variants (n=100 per model). The near-overlap of colors/sizes across panels indicates that prompt family only weakly perturbs HR relative to model-intrinsic PS/MV.
Figure 7 Prompt sensitivity–model variability facets with shared HR encoding. Each panel corresponds to a prompting technique (Zero-shot, Few-shot, CoT, Instruction, Vague). Model locations are fixed across panels by Prompt Sensitivity (PS, x-axis) and Model Variability (MV, y-axis); dashed lines mark the median cutoffs (PS = 0.079, MV = 0.057). Hallucination Rate (HR, %) is encoded by marker size (size ∝ HR) and a shared colorbar (identical range in all panels). Values are Mean ± SD over three seeds × five prompt variants (n=100 per model). The near-overlap of colors/sizes across panels indicates that prompt family only weakly perturbs HR relative to model-intrinsic PS/MV.CC BY · PubMed Central

Verification

Claims the paper makes

The claims and statistics of this paper haven’t been checked yet.

Free: it reads the open-access text and checks the first 5 claims that cite a source we can check. Create a free account to check every claim.

A few free checks a day · papers already verified open instantly · results are public

Claims other papers make about this one

Each is checked against this paper’s own text: does it say what the citing paper says it does?

Self-check

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

The self-check needs the paper’s full text.

Citations

Cites

Show all 40 references

Research works printed but not linked: 3

  1. Kazemi, 2023, Cohs: a dataset for evaluating factual consistency of summaries, arXiv [preprint], 10.4850/arXiv:2305.00038 · doi:10.4850/arxiv:2305.00038
  2. Nakano, 2021, Webgpt: browser-assisted question-answering with human feedback, arXiv [preprint], 10.48550/arXiv:2112.09332 · doi:10.48550/arxiv:2112.09332
  3. Zhang, 2023, Grounded language model training reduces hallucination, arXiv [preprint], 10.48550/arXiv:.303.08239 · doi:10.48550/arxiv:.303.08239

Cited by

See all 147

Metadata

Authors
Anh-Hoang, D., Tran, V., & Nguyen, L.-M.
Affiliations
Japan Advanced Institute of Science and Technology (JP)
Corresponding author
Anh-Hoang, Dang
Venue
Frontiers in Artificial Intelligence
Published
2025-09-30
Volume
8
Pages
1622292-1622292
ISSN
2624-8212 (ISSN-L)
OpenAlex
W4414677042
Cited by
146
References
27
Citations by year
2026: 140 · 2025: 5
FWCI
13.91
Type
journal article
Language
en
Open access
Locations
  • Frontiers in Artificial Intelligence (publishedVersion, cc-by)
  • Europe PMC (PubMed Central) (publishedVersion, cc-by)
  • PubMed (publishedVersion)
  • DOAJ (DOAJ: Directory of Open Access Journals) (submittedVersion, cc-by-sa)
Funders
Japan Society for the Promotion of Science (23K16954)
Topics
  • Physical Sciences › Computer Science › Artificial Intelligence › Topic Modeling
  • Physical Sciences › Computer Science › Artificial Intelligence › Natural Language Processing Techniques
  • Physical Sciences › Computer Science › Computer Vision and Pattern Recognition › Multimodal Machine Learning Applications
Keywords
hallucinations, large language models, prompting strategies, prompt sensitivity, model variability, chain-of-thought prompting, factual accuracy, hallucination mitigation, prompt engineering, TruthfulQA

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.

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: Survey and analysis of hallucinations in large language models: attribution to prompting strategies or model behavior. https://app.cochat.ai/papers/10.3389/frai.2025.1622292
The paper
Anh-Hoang, D., Tran, V., & Nguyen, L.-M. (2025). Survey and analysis of hallucinations in large language models: attribution to prompting strategies or model behavior. Frontiers in Artificial Intelligence. https://doi.org/10.3389/frai.2025.1622292

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.