From Twisted Facts to Rogue Agents: Can LLMs Reliably Detect Multi-Dimensional Hallucinations?

Hanyu Qiu; Ze Liu; Zhe Zhao; Kuang Xiong; Blaze Chen; Zhenxiang Yan; Renfen Hu

cs.CLNov 5, 2026

From Twisted Facts to Rogue Agents: Can LLMs Reliably Detect Multi-Dimensional Hallucinations?

Authors: Hanyu Qiu, Ze Liu, Zhe Zhao, Kuang Xiong, Blaze Chen, Zhenxiang Yan, Renfen Hu

Organizations: Beijing Normal University · Tencent AI Lab

Abstract

Large language models (LLMs) are widely used for information access, content production, and agent-based task execution, but their outputs remain vulnerable to subtle, high-risk hallucinations. To systematically evaluate and advance post-generation hallucination detection, we introduce HaluCheck, a large-scale, bilingual benchmark comprising over 42k instances spanning both factual hallucinations in text and faithfulness hallucinations in agent trajectories. HaluCheck features fine-grained, multi-level annotations for detection, localization, and correction, covering different tasks and perturbation degrees. Experiments on fourteen LLMs reveal that, while models show basic sample-level detection ability, they struggle with fine-grained localization, correction, and agent hallucination detection. Further analysis demonstrates that thinking, web search, and supervised fine-tuning each offer partial improvements, but none fully solve the problem. HaluCheck thus provides a solid basis for benchmarking and diagnosing LLM reliability across text and agent scenarios.

Figures & tables

Figure 1: Representative hallucination risks in real-world LLM applications.

Figure 2: Overview of HaluCheck . We construct a bilingual, multi-level benchmark covering factuality hallucinations in text and faithfulness hallucinations in agent trajectories. By injecting controllable perturbations into Wikipedia/QA passages and real agent interactions, HaluCheck supports hallucination detection, span localization and correction, and root-turn tracing in agent trajectories.

Dataset	Split	Lang.	Total	Obvious	Moderate	Subtle	Composition
	Full	EN	18,024	2,976	7,205	7,843	QA: 10,630; Wiki: 7,394
HaluReal	Full	CH	8,922	2,294	3,433	3,195	QA: 5,597; Wiki: 3,325
	Lite	CH/EN	1,000 / 1,000	Balanced			–
HaluFake	–	CH/EN	500 / 500	–	–	–	Fabricated entities
	Full	EN	6,743	2,206	2,286	2,251	Plan: 2,500; Perception: 2,305; Tool-use: 1,938
HaluAgent	Full	CH	8,291	2,898	2,647	2,746	Plan: 2,810; Perception: 2,691; Tool-use: 2,790

Table 1: Statistics of HaluCheck . The full sets contain balanced pairs of factual and hallucinated samples, while lite sets are used for evaluation. Perturbation-degree and composition statistics cover only hallucinated samples.

Figure 3: Hallucination rates across different entities.

Benchmark	Size	Lang.	Main focus	Data source	Evaluation protocol	Agent
TruthfulQA ( Lin et al., 2022 )	817 questions	EN	Truthfulness in QA	Misconception questions	Truthfulness and informativeness	–
HotPotQA ( Yang et al., 2018 )	113K QA pairs	EN	Multi-hop QA	Wiki QA with supporting facts	Answer/supporting-fact EM/F1	–
HaluEval ( Li et al., 2023b )	35K samples	EN	QA, dialogue, and summarization hallucination	Human-annotated LLM samples	Binary hallucination recognition	–
RAGTruth ( Niu et al., 2024 )	$\sim$ 18K responses	EN	RAG hallucination	Natural RAG responses	Case- and word-level labels	–
HALoGEN ( Ravichander et al., 2025 )	10,923 prompts; $\sim$ 150K gen.	EN	Generation hallucination measurement	LLM responses across 9 domains	Atomic-unit verification	–
AgentHallu ( Liu et al., 2026 )	693 trajectories	EN	Agent hallucination attribution	Annotated agent trajectories	Binary label, root step, explanation	Yes

Table 2: Comparison between HaluCheck and representative benchmarks.

Model	CH									EN
	HaluReal				HaluFake		HaluAgent			HaluReal				HaluFake		HaluAgent
	Acc	UR	SpanF1	CR	UR	Recall	Acc	UR	RTAcc	Acc	UR	SpanF1	CR	UR	Recall	Acc	UR	RTAcc
Gemini3.1-Pro	84.35	9.40	70.07	82.27	44.60	75.70	64.55	2.20	54.93	84.87	16.63	72.95	86.00	93.20	53.40	72.28	1.35	76.16
Gemini3-Flash	77.62	0.25	59.80	72.60	1.20	96.20	50.60	0.00	54.02	81.53	0.85	60.12	72.93	12.20	93.90	72.12	0.05	72.76
GPT5.4	63.08	16.85	39.32	56.42	60.00	67.80	50.45	0.00	57.09	71.33	15.75	42.88	63.83	92.00	54.00	72.36	0.65	73.31
GPT5-Mini	63.35	43.40	43.55	57.43	7.40	94.10	50.90	0.30	58.40	65.28	55.55	42.61	57.16	6.60	96.70	70.73	0.85	70.24

Table 3: Main results on HaluCheck . All values are percentages. UR is short for UncertainRate, CR is short for CorrectionRate, and RTAcc is short for RootTurnAccuracy. The best result for each metric is highlighted in bold .

Component	Case 1	Case 2	Case 3
Input Snippet	“… Samsung TV Plus … due to its integration with the webOS operating system.”	“The 2014 crime drama film starring Tom Hardy and James Gandolfini … was directed by Michaël R. Roskam.”	“He is currently Professor … at McGill University, and Recurrent Visiting professor … at the Central European University in Prague, Hungary.”
GroundTruth	Hallu : Tizen $\rightarrow$ webOS	Clean : no hallucination	Hallu : Queen’s University $\rightarrow$ McGill University; Budapest, Hungary $\rightarrow$ Prague, Hungary
Key CoT Content	“ webOS is actually used by LG , not Samsung. Samsung uses Tizen OS for their smart TVs.”	“ James Gandolfini died in 2013 , but the film was released in 2014 . However, the film was shot before his death and released afterward, so he could still appear in it .”	“ Kymlicka is associated with Queen’s University , not McGill. Also, Central European University is in Budapest , Hungary, not ‘Prague, Hungary’.”

Table 4: Representative cases where enabling thinking corrects hallucination detection decisions.

Figure 4: Effect of thinking mode (Mean

\pm

std. across five Qwen3 Models).

Figure 5: Effect of web search (Mean

\pm

std. across five Qwen3 Models).

Figure 6: Effect of LoRA Fine-tuning.

Appendix figures & tables10 assets

Supplementary material from the paper’s appendix.

Appendix

Language	Hallucinated	Clean
CH	16.74	16.67
EN	12.14	11.05

Appendix

Table 5: Qwen3-8B-Base perplexity comparison between hallucinated and clean texts.

Metric	Rule	LLM	$r$	$\rho$
Span-P	42.9	34.9	0.965	0.915
Span-R	48.6	41.5	0.975	0.964
Span-F1	44.4	36.8	0.967	0.948
CorrRate	62.6	47.1	0.645	0.635

Appendix

Table 8: Correlation between rule-based span matching and LLM-judge evaluation on HaluReal .

Metric	$r$	$\rho$	Qwen3.5	GLM4.7
Precision	0.9948	0.9895	0.3520	0.3089
Recall	0.9704	0.9862	0.4238	0.3523
F1	0.9892	0.9886	0.3737	0.3208
Correction Acc.	0.9787	0.9780	0.4740	0.4311

Appendix

Table 9: Robustness of span-level evaluation across Qwen3.5 and GLM4.7 judges.

Figure 7: Average model performance across perturbation degrees on HaluReal/Text and HaluAgent. Error bars denote one standard deviation across models.

Figure 8: Average HaluAgent performance across perturbation categories. Error bars denote one standard deviation across models.

Item	Value
Framework / optimizer	ms-swift ( swift sft ) + DeepSpeed ZeRO-2 / AdamW
Tuner	LoRA: rank 16, alpha 32, dropout 0.05
LoRA modules	all-linear; bias none
Precision	bf16; gradient checkpointing
Per-device batch	train 2 / eval 2
Gradient accumulation	4

Appendix

Table 10: Hyperparameters for LoRA-SFT experiments.

Dataset Pair	Metric	Pearson	Spearman
HaluText -CH	Accuracy	$0.8722^{***}$	$0.8273^{***}$
	Precision	$0.8376^{***}$	$0.8889^{***}$
	Recall	$0.9069^{***}$	$0.8304^{***}$
	F1	$0.9177^{***}$	$0.9077^{***}$
	Refusal Rate	$0.9536^{***}$	$0.9063^{***}$
HaluText -EN	Accuracy	$0.9499^{***}$	$0.9253^{***}$

Appendix

Table 11: Model-level correlations between HaluCheck and natural-hallucination test sets. ∗ , ∗∗ , and ∗∗∗ denote significance at

p<0.05

p<0.01

, and

p<0.001

, respectively.

Figure 9: FactScore vs SpanF1

Panel of Figure 10 — Figure 10: Prompt used for HaluText

Panel of Figure 11 — Figure 11: Prompt used for HaluAgent

Hallucination in Language Models

Hallucination in language models studies how systems produce unsupported or false claims and how to detect, explain, and reduce them.

See all 261 papers

HalluScan: A Systematic Benchmark for Detecting and Mitigating Hallucinations in Instruction-Following LLMs
May 4, 2026Ahmed CherifLLM Evaluation LLM Hallucination Mitigation
1Sofrecom Tunisia, Orange Innovation, Tunis, 1053, Tunisia.
OpenHalDet: A Unified Benchmark for Hallucination Detection across Diverse Generation Scenarios
Jun 5, 2026Xinyi Li, Zhen Fang, Yongxin Deng +12LLM Evaluation Hallucination Detection
University of Technology Sydney · University of Bristol · University of Wisconsin–Madison +2
PARALLAX: Separating Genuine Hallucination Detection from Benchmark Construction Artifacts
May 16, 2026Khizar Hussain, Murat KantarciogluLLM Evaluation Language Model Error Detection
Virginia Tech

May 4, 2026J/K move · Enter open · S save

HalluScan: A Systematic Benchmark for Detecting and Mitigating Hallucinations in Instruction-Following LLMs

Ahmed Cherif

1Sofrecom Tunisia, Orange Innovation, Tunis, 1053, Tunisia.

LLM Evaluation LLM Hallucination Mitigation Hallucination in Language Models LLM Reliability LLM Hallucination Detection

Open paper AskPDF arXiv

Large Language Models (LLMs) have demonstrated remarkable capabilities across diverse natural language processing tasks, yet they remain susceptible to hallucinations -- generating content that is factually incorrect, unfaithful to provided context, or misaligned with user instructions. We present HalluScan, a comprehensive benchmark framework that systematically evaluates hallucination detection and mitigation across 72 configurations spanning 6 detection methods, 4 open-weight model families, and 3 diverse domains. We introduce three key contributions: (1) HalluScore, a novel composite metric that achieves a Pearson correlation of r = 0.41 with human expert judgments; (2) Adaptive Detection Routing (ADR), an intelligent routing algorithm achieving 2.0x cost reduction with only 0.1% AUROC degradation; and (3) systematic error cascade decomposition revealing substantial variation in hallucination error types across domains. Our experiments reveal that NLI Verification achieves the highest overall AUROC of 0.88, while RAV achieves the second-highest AUROC of 0.66.

Abstract

Hallucination in Language Models

Explore similar work

HalluScan: A Systematic Benchmark for Detecting and Mitigating Hallucinations in Instruction-Following LLMs

OpenHalDet: A Unified Benchmark for Hallucination Detection across Diverse Generation Scenarios

PARALLAX: Separating Genuine Hallucination Detection from Benchmark Construction Artifacts