When Models Lie, We Learn: Multilingual Span-Level Hallucination Detection with PsiloQA

Introduction

• LLMs can produce unsupported facts, and even one hallucinated word can alter meaning and cause harm, making detection critical in safety-sensitive uses (Xiao and Wang, 2021; Xu et al., 2024).The passage cites medicine as an example of a safety-critical domain (Ben Abacha and Demner-Fushman, 2019; He et al., 2025).Source: §1 · p. 1
• Sampling-based UQ methods incur substantial overhead and operate at sequence level, while information-based methods can miss span hallucinations; retrieval-based fact-checking also costs more and depends on evidence and auxiliary-LLM quality.The cited approaches include sampling-based methods (Duan et al., 2024), information-based methods (Fomicheva et al., 2020; Fadeeva et al., 2024), and external-knowledge or auxiliary-LLM fact-checking (Niu et al., 2024; Mishra et al., 2024).Source: §1 · p. 2
• Span-level annotations are labor-intensive, and multilingual data creation adds challenges from linguistic diversity, limited non-English data, and language-specific pipeline adjustments (Vashurin et al., 2025).The passage notes that span-level annotation is more complex than sequence-level verification.Source: §1 · p. 2
• The paper introduces an automated four-step pipeline that generates multilingual Wikipedia question-answer pairs, elicits no-context answers, annotates inconsistent spans, and filters unsuitable cases.The figure specifies GPT-4o for question-answer generation and span-level inconsistency detection.Source: §1 · p. 2, Figure 1 · p. 2
• PsiloQA is a large multilingual dataset with fine-grained span-level hallucination annotations for answers from open-weight and proprietary LLMs.The paper also reports comprehensive evaluations of state-of-the-art detection methods across 14 languages.Source: §1 · p. 2

2.1 Hallucination Detection Datasets

• Sentence- or paragraph-level benchmarks such as TruthfulQA (Lin et al., 2022), ANAH (Ji et al., 2024; Gu et al., 2024), and HaluEval (Li et al., 2023) cannot locate hallucinated content within responses.This is particularly limiting for long-form text containing both supported and unsupported information, where a binary response label is insufficient (Min et al., 2023).Source: §2.1 · p. 3
• FActScore (Min et al., 2023) supports fine-grained detection in Wikipedia bibliographies, while HalluEntity (Yeh et al., 2025) labels atomic facts in ChatGPT biographies as true or false.FActScore supports evaluation of methods using external knowledge and fact-level annotations; HalluEntity aligns its labels with the relevant entity in the generated biography.Source: §2.1 · p. 3
• RAGTruth (Niu et al., 2024) provides nearly 18,000 human-annotated, word-level hallucination spans for RAG tasks, while FAVA (Mishra et al., 2024) supports fine-grained detection and editing across generation tasks.RAGTruth covers question answering, data-to-text, and summarization; FAVA uses external knowledge support.Source: §2.1 · p. 3
• Mu-SHROOM targets hallucination-span detection in instruction-tuned LLM outputs across Arabic, Basque, Catalan, Chinese, Czech, English, Farsi, Finnish, French, German, Hindi, Italian, Spanish, and Swedish.The task is described as part of the shared task on Multilingual Hallucinations and Related Observable Overgeneration Mistakes.Source: §2.1 · p. 3

2.2 Hallucination Detection Methods

• Detection methods include sampling-based uncertainty estimation, prompted self-evaluation, and supervised approaches that operate at sequence or token level.Supervised approaches often require model-specific training because models differ in hidden-feature and attention-head counts.Source: §2.2 · p. 3
• Span-level methods use token probabilities or entropy, compare probable token candidates with NLI through Claim Conditioned Probability (Fadeeva et al., 2024), or reweight uncertainty using prior-token uncertainty and attention (Zhang et al., 2023).Sampling-based uncertainty methods measure consistency across sampled generations (Kuhn et al., 2023; Lin et al., 2023; Duan et al., 2024; Zhang et al., 2024).Source: §2.2 · pp. 3–4
• Knowledge-based methods include FActScore (Min et al., 2023), which checks extracted atomic facts against retrieved context, and RAGTruth (Niu et al., 2024), which benchmarks human-annotated word-level hallucinations in RAG responses.FAVA (Mishra et al., 2024) extends detection to word-level hallucination editing using a taxonomy and LLM-generated errors; RAGTruth and FAVA are limited to English and require human annotations.Source: §2.2 · p. 4

3.1 Dataset Generation Process

• PsiloQA builds multilingual Wikipedia QA pairs, elicits no-context answers from Mu-SHROOM models, then uses GPT-4o to mark discrepant word-level hallucination spans.Rule- and prompt-based filters remove malformed annotations, subjective or incomplete questions, and refusals; 6,575 samples were filtered. GPT-4o annotation is described as scalable, faster, and more cost-effective than human annotation.Source: §3.1 · pp. 4–5

3.2 Dataset Statistics

• English is the largest language group at nearly 23,000 samples and German the smallest at about 1,500; Geography and Sports each comprise roughly 25–30% of samples.There are 14,000 samples with no hallucinations and 50,000 with one span; 50,000 spans contain fewer than five words.Source: §3.2 · p. 5

3.3 Dataset Production Cost

• GPT-4o labeling was estimated at $535, compared with a rough estimate of approximately $3,000 for RAGTruthQA span labeling.RAGTruth responses were labeled by two annotators, with reported consistency of 91.8% at response level and 78.8% at span level.Source: §3.3 · p. 5

3.4 Manual Dataset Analysis

• GPT-4o’s annotations achieved 84.3% AP and 71.0% IoU against aggregated human labels, while annotators agreed at 80.1% AP and 76.8% IoU.The study used 100 randomly selected English test samples; its reported worst-case margin of error was approximately 9.8% at the 95% confidence level.Source: §3.4 · p. 5

4.1 Datasets

• The evaluation uses PsiloQA and four QA-based benchmarks for span-level hallucination detection.The benchmarks are Mu-SHROOM, FAVA-Bench, HalluEntity, and RAGTruthQA.Source: §4.1 · p. 6
• Mu-SHROOM covers 14 languages; its description reports 3,351 unlabeled samples in English, Spanish, Chinese, and French, and a 1,902-sample test set across eight languages.The test set has around 100 items each for Basque, Catalan, Czech, and Farsi, and around 150 for other languages.Source: §4.1 · p. 6
• FAVA-Bench is an English human-annotated span-level benchmark; HalluEntity has 157 English human-annotated samples; RAGTruthQA has 900 English human-annotated samples.FAVA-Bench focuses on span-level detection because most models could not detect errors using its taxonomy; HalluEntity uses generated biographies, and RAGTruthQA uses MS MARCO questions.Source: §4.1 · p. 6

4.2 Metrics

• Evaluation combines span-level intersection over union (IoU) with character-level average precision (AP), which is threshold-independent and suited to imbalanced data and cross-dataset transfer.Span annotations are binarized at character level for IoU; AP measures area under the precision-recall curve.Source: §4.2 · p. 6

4.3.1 Uncertainty Quantification

• Token-level UQ baselines are MaxProb, CCP, and Focus, with language-specific validation-set thresholds used to compute IoU.These methods are applied to each token because many UQ approaches operate at sequence level or require model-specific training.Source: §4.3.1 · p. 6

4.3.2 Encoder Models

• Encoder methods include RAGTruth-trained LettuceDetect, multilingual mmBERT-base, and ModernBERT-base fine-tuned on PsiloQA using LettuceDetect's hyperparameters.LettuceDetect uses ModernBERT and supports 8,192-token contexts, but its English-focused pretraining limits suitability for other languages; mmBERT-base provides native multilingual support.Source: §4.3.2 · p. 7

4.3.3 Language Models

• The LLM approaches are GPT-4o-based FActScore and 3-shot Qwen2.5-32B-Instruct prompting; FActScore marks tokens occurring in unsupported claims using a 0.5 threshold.Tokens in all claims or present in the original input are excluded from hallucination marking; Qwen examples are randomly selected from the validation set.Source: §4.3.3 · p. 7

5.1 Performance on PsiloQA

• Among uncertainty methods, Focus has the highest AP scores, while FActScore and 3-shot Qwen2.5-32B-it show language-specific AP strengths but weaker or moderate span-boundary IoU.Focus’s example AP scores are 68.90 for Finnish and 63.61 for English; FActScore has poor IoU, and Qwen’s IoU remains moderate.Source: §5.1 · pp. 7–8
• mmBERT achieves the highest scores in 12 of 14 languages for both metrics, and fine-tuned multilingual encoders outperform uncertainty-based and LLM-based methods overall.The gap between AP and IoU across methods indicates that precise span-boundary detection remains challenging.Source: §5.1 · pp. 7–8

5.2 Cross-lingual Transfer

• Training mmBERT on multilingual PsiloQA generally outperforms language-specific training across PsiloQA and Mu-SHROOM test languages, including Arabic and Hindi.The reported gains extend across languages with distinct scripts and different language families.Source: §5.2 · p. 8, Table 3 · p. 8

5.3 Knowledge Transfer

• PsiloQAen-trained models outperform RAGTruthQA across benchmarks, including a 45% IoU improvement on Mu-SHROOMen; joint training yields the highest AP on HalluEntity.On Mu-SHROOMen, the combined model ranks second to PsiloQA alone, while FAVA-Bench shows limited performance across configurations.Source: Table 4 · p. 8, §5.3 · p. 9
• PsiloQA’s advantage over RAGTruthQA may reflect its larger training set, while synthetic generation costs more than 17 times less than manually curated RAGTruthQA.The passage also characterizes PsiloQA’s generation cost as significantly lower than RAGTruth’s labeling cost.Source: §5.3 · p. 9

Conclusion

• PsiloQA is a large-scale multilingual span-level dataset built from real zero-context LLM hallucinations, with GPT-4o automated span annotations, broad language coverage, and support for diverse LLM architectures.The dataset is designed as a resource for evaluating and training hallucination detection models.Source: Conclusion · p. 9
• Fine-tuned encoder methods, especially mmBERT and ModernBERT, outperform other baselines; PsiloQA generalizes cross-lingually and outperforms RAGTruth in knowledge transfer despite being over 17 times cheaper to produce.Source: Conclusion · p. 9
• Evaluations across uncertainty quantification, encoder-based detectors, and LLM-based methods demonstrate PsiloQA's effectiveness as a benchmark, although significant hallucination-detection challenges remain.Source: Conclusion · p. 9
• Future work will extend the PsiloQA pipeline to other generation tasks, including summarization and data-to-text generation, to broaden its utility in hallucination research.Source: Conclusion · p. 9

Limitations

• GPT-4o generates and annotates PsiloQA, risking model-specific bias and cross-lingual inconsistency; the authors propose multiple annotators and human-in-the-loop validation.GPT-4o judgments may not reflect human consensus, particularly in edge cases; an ensemble with span averaging is suggested for future work.Source: Limitations · p. 9, Ethical Considerations · p. 10
• PsiloQA covers only question answering and does not explicitly represent a diverse taxonomy of hallucination types, potentially skewing error-type coverage.Its hallucinations arise naturally from LLM errors in a zero-context setting; summarization, dialogue, and data-to-text remain untreated tasks.Source: Limitations · p. 9
• Uneven language resources and Wikipedia coverage may introduce cultural or regional bias, while limiting topical diversity and fit to noisy, domain-specific applications.Lower-resource languages may have fewer high-quality examples, and many baselines are optimized predominantly for English, potentially underestimating performance in other languages.Source: Limitations · pp. 9–10

Ethical Considerations

• PsiloQA is intended for research, not high-stakes or real-time deployment without domain-specific validation; the authors discourage surveillance, censorship, and automated moderation without human oversight.The dataset contains no personally identifiable information or sensitive or private data, and its documentation includes the generation and annotation pipeline, prompts, and filtering mechanisms.Source: Ethical Considerations · p. 10
• The dataset spans 14 languages, but uneven language coverage and GPT-4o’s varying proficiency may cause performance disparities and inconsistent cross-lingual annotations.The authors identify diverse model perspectives and human-in-the-loop validation as possible mitigations.Source: Ethical Considerations · p. 10

A License and Infrastructure

• Experiments used 2 NVIDIA A100 GPUs for around 50 GPU-hours; Qwen 2.5 is under Apache 2.0, and PsiloQA is released under CC-BY-4.0.Source: A License and Infrastructure · p. 13

B Packages

• Dataset generation used VLLM (Kwon et al., 2023) with default Hugging Face README hyperparameters, while encoder training used Transformers.Training hyperparameters are detailed in Appendix C.Source: B Packages · p. 13

C Encoder Hyperparameters

• Training followed LettuceDetect (Kovács and Recski, 2025) for six epochs, selecting the checkpoint with the best validation performance.Source: Table 5 · p. 13

• Table 6 reports the utilized LLMs and their languages, average number of spans, and average span length.Source: Table 6 · p. 14
• Figure 2 shows the number of PsiloQA samples generated by each model.Source: Figure 2 · p. 14
• Figure 3 presents the distribution of LLM answer lengths in tokens, measured using spaCy.Source: Figure 3 · p. 15
• Figure 5 shows the distribution of hallucination spans in PsiloQA.Source: Figure 5 · p. 15
• Figure 4 presents the distribution of languages in the PsiloQA dataset.Source: Figure 4 · p. 15
• Figure 6 presents the distribution of hallucination span lengths in tokens, measured using spaCy.Source: Figure 6 · p. 15

E PsiloQA Dataset Sample

• Each PsiloQA sample includes a retrieved Wikipedia passage, a GPT-4o-generated question and golden answer, an LLM answer, and annotated hallucination ranges.The golden answer is generated at a specified complexity level.Source: Table 7 · p. 16
• GPT-4o annotates hallucination ranges by comparing each LLM answer with the golden answer.Table 7 illustrates these ranges as highlighted spans.Source: Table 7 · p. 16

F QA Pairs Generation Prompt

• The QA-generation prompt asks GPT-4o to create three question-answer pairs of different complexity from a retrieved Wikipedia passage and controls answer length.In 33% of cases, it requests long, detailed answers.Source: Figure 7 · p. 17

G Inconsistency Detection Prompt

• The inconsistency-detection prompt supplies questions, LLM answers, and GPT-4o-generated golden answers, then asks GPT-4o to identify inconsistent spans.Inconsistencies with the golden answer and context are treated as hallucinations.Source: Figure 8 · p. 18

H Filtering Prompts

• The filtering prompt identifies subjective questions that call for personal opinions rather than factual answers.The prompt also detects incomplete questions lacking a clear subject and therefore unanswerable.Source: Figure 9 · p. 19
• A separate filtering prompt detects cases in which an LLM refuses to answer.Source: Figure 10 · p. 20

I LLM Baseline Evaluation Prompt

• The baseline is evaluated in 3-shot mode using PsiloQA validation examples, with prompts translated and examples selected from the corresponding language.Figure 11 shows the English prompt.Source: Figure 11 · p. 20