Hallucination stress tests can yield similar average scores across language models even when different prompts elicit hallucinations in each model. We develop a procedure that iteratively edits source-grounded seed prompts to expose hallucinations in a target model. An LLM editor revises each prompt in response to the target's answer and a judge's category-specific score and explanation. In the main setting, the target also receives feedback on its previous response. Experiments cover 500 seeds from seven datasets and ten topics, with intrinsic and extrinsic hallucinations evaluated across open-weight and closed-source models. Editing substantially increases judge-rated severity over the seeds. In the baseline configuration, retained responses fall in the judge's strong-hallucination bin in 77.2% of runs targeting extrinsic hallucination and 74.2% targeting intrinsic hallucination. Later editing rounds account for more than half of the mean score increase. Among 219 prompts that pass model-assisted source-validity checks, 89 elicit responses containing unsupported claims. Across five models, 200 curated prompts yield similar average severity scores, although pairwise Spearman correlations between prompt-level scores are only ρ=0.0387 to 0.1897. Curated prompts can be reused across models, but their diagnostic value depends on source validity and the failures they expose in each model.
Figures & tables
Figure 2: Target-aware hallucination stress-test curation. Underlined text indicates the controlled hallucination category. Text in red is provided to agents after the first editing round.
Topic
Source Dataset
Amount
Cooking
QookA
50
Crime/Justice
LegalQA
50
Entertainment/Media/Pop Culture
RAGTruth + BoolQ
50
Environment
ClimRetrieve
50
Everyday Life
RAGTruth + BoolQ
50
Health/Medicine
BoolQ + RAGTruth
50
Table 1: Source distribution of the 500 seed prompts.
Figure 3: Before- and after-editing judge scores and signed SU–VU gaps for extrinsic and intrinsic hallucinations across model combinations. QA, E, and J indicate the target QA, editor, and judge models.
Figure 4: Topic-level hallucination-induction success rates for the baseline setting with Mistral-7B as target and editor and LLaMA-8B as judge.
Category
Original prompt intent
Edited prompt fragment
Stress-test mechanism
Extrinsic
Ask about Mississippi’s role in a historical source.
“Can you elaborate on Mississippi’s secession from the Union and its position within the Confederate states?”
The edit asks for broader historical detail than the source directly supports.
Extrinsic
Ask about ingredients in a recipe with limited source information.
“What can I do with the oil and butter from the recipe … but does not specify the type of oil or butter to use?”
The edit removes the source-grounding guardrail and pressures the target to infer unsupported recipe details.
Intrinsic
Ask about climate evidence from a study.
“You are a climate scientist … asked to write an article about regional differences in atmospheric conditions … however, you only have access to a study that focuses on sea surface temperature drifts …”
The edit creates a mismatch between the requested inference and the available evidence.
Table 2: Representative edited prompts. Source passages are omitted for brevity. Nore examples in Appendix F .
Figure 5: Judge scores and signed SU–VU gaps for target-specific prompts and prompts transferred from the setting targeting Mistral-7B. Target-specific scores are shown for comparison.
Appendix figures & tables7 assets
Supplementary material from the paper’s appendix.
Appendix
Judge model
No hallucination
Weak hallucination
Strong hallucination
LLaMA-8B
0–37
38–75
76–100
LLaMA-3B
0–33
34–71
72–100
Qwen2.5-7B
0–22
23–70
71–100
OpenAI o3
0–26
27–66
67–100
Appendix
Table A1: Judge score thresholds for hallucination severity bins across four judge models.
Figure A1: Before- and after-editing judge scores and signed SU–VU gaps for extrinsic hallucination under different judge models.
Figure A2: Before- and after-editing judge scores and signed SU–VU gaps for intrinsic hallucination under different judge models.
Figure C1: MTurk questionnaire for validating hallucination scores assigned by the judge. Red questions are used for extrinsic hallucination, and purple questions are used for intrinsic hallucination.
Figure C2: MTurk questionnaire for validating hallucination explanations generated by the judge.
Figure D1: Relationship between judge hallucination severity score and signed SU–VU gap over 2,200 instances. The relationship is positive but weak, so SU–VU is used as a complementary diagnostic rather than an independent ground-truth validator.
Figure F1: Before- and after-editing judge scores and signed SU–VU gaps for extrinsic and intrinsic hallucination across three random seeds.
Large Language Models (LLMs) possess a remarkable capacity to generate persuasive and intelligible language. However, coherence does not equate to truthfulness, as the responses often contain subtle hallucinations. Existing benchmarks are limited by static and narrow questions, leading to limited coverage and misleading evaluations. We present KGHaluBench, a Knowledge Graph-based hallucination benchmark that assesses LLMs across the breadth and depth of their knowledge, providing a fairer and more comprehensive insight into LLM truthfulness. Our framework utilises the KG to dynamically construct challenging, multifaceted questions, whose difficulty is then statistically estimated to address popularity bias. Our automated verification pipeline detects abstentions and verifies the LLM's response at both conceptual and correctness levels to identify different types of hallucinations. We evaluate 25 frontier models, using novel accuracy and hallucination metrics. The results provide a more interpretable insight into the knowledge factors that cause hallucinations across different model sizes. KGHaluBench is publicly available to support future developments in hallucination mitigation.
Alex Robertson, Huizhi Liang, Mahbub Gani +2
School of Computing, Newcastle University · Sage Ai, Sage Group PLC · Current affiliation: Redis
Most hallucination evaluations focus on English, leaving it unclear whether findings transfer to lower-resource languages. We investigate faithfulness hallucinations, defined as model-generated content that is fluent and plausible but diverges from the provided input or is internally inconsistent. Leveraging the multilingual MultiWikiQA dataset, we utilize the LettuceDetect framework to create synthetic hallucination datasets for 306 languages, from which we train token-level hallucination classifiers for 30 European languages. In this work, we present evaluations of model hallucinations on a selection of languages: English, Danish, German, and Icelandic. Using these classifiers, we evaluate the hallucination rates for Qwen3-0.6B, Qwen3-14B, Gemma-3-12B-IT, cogito-v1-preview-qwen-32B, and cogito-v1-preview-llama-70B. Our classifiers reveal notably higher hallucination rates for Qwen3-0.6B (up to 60% of answers containing at least one hallucination, peaking in Icelandic) and generally lower rates for larger models, with cogito-v1-preview-qwen-32B and cogito-v1-preview-llama-70B performing best on most languages. Hallucination rates are consistently higher for lower-resource languages, particularly Icelandic.
Freja Thoresen, Dan Saattrup Smart
Alexandra Institute · Rued Langgaards Vej 7, 2300 København
Large language models (LLMs) hallucinate with confidence: their outputs can be fluent, authoritative, and simply wrong. In medical, legal, and scientific applications this failure causes direct harm, and detecting it from internal model states offers a path to safer deployment. A growing body of work reports that this problem is increasingly tractable, with recent methods achieving high detection performance on widely used benchmarks. We show, however, that much of this apparent progress does not survive scrutiny. Four of the six corpora embed the ground-truth answer directly in the input prompt. A naïve text-similarity baseline we call TxTemb exploits this to achieve near-perfect detection scores without any access to model internals. To measure what genuine detection capability remains once these artifacts are controlled, we conduct a large-scale evaluation spanning twenty-two detection methods, twelve open-source models spanning six architectural families, and six corpora. We further introduce DRIFT, a supervised probe over inter-layer hidden-state transitions, as a point of comparison for live-generation detection. Our findings suggest that the field's reported progress on hallucination detection is substantially explained by benchmark construction artifacts in widely used corpora, and that the majority of established baselines perform near chance under controlled conditions; the consistent exceptions are SAPLMA and DRIFT, both supervised probes on upper-layer hidden states.