UZH at SemEval-2025 Task 3: Token-Level Self-Consistency for Hallucination Detection

Michelle Wastl, Jannis Vamvas, Rico SennrichACL 2025

TL;DR

This paper introduces a novel token-level self-consistency approach for detecting hallucinations in multilingual text generated by Large Language Models (LLMs). By analyzing the consistency of generated text at the sub-word level, their method achieves competitive results on the SemEval-2025 Task 3, outperforming simple baselines and showing promise even for low-resource languages.

What Problem Does This Paper Address?

Large Language Models (LLMs) are increasingly powerful, but their outputs can contain "hallucinations" – fabricated or factually incorrect information. Detecting these hallucinations is crucial for deploying LLMs reliably in real-world applications. While previous work has focused on sentence-level detection or relied on external knowledge, this paper tackles the challenge of identifying hallucinated spans of text within LLM outputs across multiple languages, including low-resource ones. The core problem is that hallucinations often occur at a fine-grained, sub-sentential level, making sentence-level detection insufficient.

Methods

The authors propose two main consistency-based approaches:

  • Token-level self-consistency: This method evaluates the consistency of the generated text by comparing it to multiple alternative responses generated by the same underlying LLM. The idea is that if a specific span of text is consistently generated across different sampling configurations of the same model, it's more likely to be factual.
  • Token-level cross-model consistency (GPT-consistency): This approach uses a superior LLM (GPT-4o-mini) to compare the target text against alternative responses. It aligns tokens from the text to be evaluated with tokens from the alternative answers, using the median similarity score as an indicator of consistency.

Both methods operate at the token level, allowing for more granular detection than previous sentence-level approaches. They also introduce a word-level label derivation by majority voting from subword token labels.

Datasets

The research utilizes data from the SemEval-2025 Task 3: MuSHROOM dataset. This dataset includes question-answer pairs with human-annotated hallucinated spans for 10 languages (Arabic, Chinese, English, French, Finnish, German, Hindi, Italian, Spanish, Swedish) and unlabelled data for 4 additional languages (Farsi, Basque, Czech, Catalan). The authors split the validation set into training and validation sets for their experiments.

Key Insights and Ideas

  • Token-level consistency is key: Hallucinations often occur at a sub-sentential level, making token-level analysis crucial for accurate detection.
  • Self-consistency is model-dependent: The effectiveness of self-consistency is highly sensitive to the specific underlying model and its size. Larger models (>8B parameters) benefit more from self-consistency compared to smaller ones.
  • GPT-consistency offers stability: Using a superior LLM for cross-model consistency appears to be more stable across different languages and models, showing less performance degradation.
  • Threshold calibration matters: Accurate threshold calibration, especially for each model and language, is vital for the success of consistency-based approaches.
  • Prompting can help, but consistency is competitive: While prompt engineering for zero-shot LLMs can yield strong results, consistency-based methods can be competitive, especially for certain languages and models.

Key Results

  • The proposed token-level consistency approaches outperform simple "mark-all" baselines.
  • Their methods achieve competitive results compared to other submissions in the shared task.
  • For some languages, their GPT-consistency approach rivals GPT-4o-mini prompt-based methods.
  • For languages where the underlying models were larger (>8B parameters), self-consistency consistently outperformed the zero-shot baseline.
  • Performance degradation for previously unseen languages was observed, particularly for self-consistency, highlighting the importance of model- and language-specific threshold calibration.

Limitations and Future Work

  • Computational Cost: Consistency-based approaches require generating multiple alternative responses, which can be computationally expensive and time-consuming. The paper explores reducing the number of alternative responses to mitigate this.
  • Model and Language Specificity: The effectiveness of self-consistency is highly dependent on the specific model and language, requiring careful threshold calibration. GPT-consistency shows more stability.
  • Small Sample Sizes: The analysis of model-specific performance was limited by small and imbalanced sample sizes, suggesting that experiments would need to be repeated on a larger dataset for more robust conclusions.
  • Future Work: The authors suggest exploring ways to improve stability across smaller models by combining self-consistency and GPT-consistency. They also propose extending these consistency-based methods to other generative tasks beyond question-answering.

Summary updated

BibTeX
@inproceedings{wastl2025uzh,
  author = {Michelle Wastl and Jannis Vamvas and Rico Sennrich},
  title = {{UZH at SemEval-2025 Task 3: Token-Level Self-Consistency for Hallucination Detection}},
  booktitle = {Annual Meeting of the Association for Computational Linguistics},
  year = {2025},
}
Figure