Can LLMs Detect Intrinsic Hallucinations in Paraphrasing and Machine Translation?

Evangelia Gogoulou, Shorouq Zahra, Liane Guillou, Luise Dürlich, Joakim Nivre

Introduction

As Large Language Models (LLMs) are increasingly deployed in critical workflows, the issue of "hallucinations"—outputs that are fluent but factually incorrect—remains a significant barrier. This study focuses on intrinsic hallucinations, where the generated text contradicts the provided source material, specifically in paraphrasing and machine translation tasks. The research evaluates whether open-access LLMs can effectively serve as their own evaluators in a contrastive detection setup.

Methodology

The researchers evaluated several model families, including Llama-3, Mixtral, EuroLLM, and GPT-SW3, in a zero-shot setting. A key component of the study was prompt engineering, testing six distinct formulations to determine if explicit terminology like "hallucination" or formal definitions influenced detection accuracy.

Table 1: The six prompt variations used to evaluate model sensitivity to task framing.

The study utilized the HalluciGen dataset, covering English and Swedish paraphrasing, as well as bidirectional translation for English-German and English-French. Hallucinations were categorized into eleven types, such as named-entity swaps, negation flips, and numerical errors. Models were prompted in both English and the target language to assess cross-lingual robustness.

Results and Discussion

The experimental results revealed several counter-intuitive findings regarding model scale and specialization.

The NLI Advantage

Specialized Natural Language Inference (NLI) models, such as BGE-M3, emerged as highly effective baselines. In paraphrasing tasks, these smaller models achieved F1 scores near 0.90, frequently outperforming much larger LLMs. This suggests that for tasks requiring strict logical entailment, massive parameter counts are not always necessary.

Table 2: Performance comparison on English and Swedish paraphrase detection across various LLMs and NLI baselines.

Model Scaling and Instruction Tuning

While Llama-3-70B-Instruct and Mixtral-8x7B-Instruct were top performers, scaling did not always correlate with better detection. Notably, Mixtral-8x22B performed significantly worse than the smaller Mixtral-8x7B, often producing "invalid" responses rather than incorrect ones. Instruction tuning generally benefited larger models but had mixed results for smaller, specialized models like GPT-SW3.

Prompting and Language Effects

Explicitly using the term "hallucination" in prompts often hindered performance; models generally fared better when asked to identify "logical contradictions" or "bad translations." Furthermore, English prompting consistently outperformed native-language prompting (e.g., Swedish or French), likely due to the English-centric nature of the models' pre-training data.

Conclusion

While top-tier LLMs show promise in detecting intrinsic hallucinations, they are often matched or exceeded by specialized NLI models in paraphrasing tasks. LLMs maintain a slight edge in machine translation, likely due to their broader cross-lingual capabilities. Future work should explore hybrid systems that leverage LLMs to generate synthetic training data for more efficient, NLI-based detection tools.

Summary updated

BibTeX
@inproceedings{gogoulou2025can,
  author = {Evangelia Gogoulou and Shorouq Zahra and Liane Guillou and Luise Dürlich and Joakim Nivre},
  title = {{Can LLMs Detect Intrinsic Hallucinations in Paraphrasing and Machine Translation?}},
  booktitle = {Annual Meeting of the Association for Computational Linguistics},
  year = {2025},
}
Figure