ML Anthology

ECD: Efficient Contrastive Decoding with Probabilistic Hallucination Detection

Laura Fieback, Nishilkumar Balar, Jakob Spiegelberg, Hanno Gottschalk
ECML-PKDD 2025 pp. 21-38
doi:10.1007/978-3-032-06109-6_2 /ecmlpkdd/2025/fieback2025ecmlpkdd-ecd/

Abstract

Despite recent advances in Large Vision Language Models (LVLMs), these models still suffer from generating hallucinatory responses that do not align with the visual input provided. To mitigate such hallucinations, we introduce Efficient Contrastive Decoding (ECD), a simple method that leverages probabilistic hallucination detection to shift the output distribution towards contextually accurate answers at inference time. By contrasting token probabilities and hallucination scores, ECD subtracts hallucinated concepts from the original distribution, effectively suppressing hallucinations. Notably, our proposed method can be applied to any open-source LVLM and does not require additional LVLM training. We evaluate our method on several benchmark datasets and across different LVLMs. Our experiments show that ECD effectively mitigates hallucinations, outperforming state-of-the-art methods with respect to performance on LVLM benchmarks and computation time.

PDF ECML-PKDD Semantic Scholar

Cite

Text

Fieback et al. "ECD: Efficient Contrastive Decoding with Probabilistic Hallucination Detection." European Conference on Machine Learning and Principles and Practice of Knowledge Discovery in Databases, 2025. doi:10.1007/978-3-032-06109-6_2

Markdown

[Fieback et al. "ECD: Efficient Contrastive Decoding with Probabilistic Hallucination Detection." European Conference on Machine Learning and Principles and Practice of Knowledge Discovery in Databases, 2025.](https://mlanthology.org/ecmlpkdd/2025/fieback2025ecmlpkdd-ecd/) doi:10.1007/978-3-032-06109-6_2

BibTeX

@inproceedings{fieback2025ecmlpkdd-ecd,
  title     = {{ECD: Efficient Contrastive Decoding with Probabilistic Hallucination Detection}},
  author    = {Fieback, Laura and Balar, Nishilkumar and Spiegelberg, Jakob and Gottschalk, Hanno},
  booktitle = {European Conference on Machine Learning and Principles and Practice of Knowledge Discovery in Databases},
  year      = {2025},
  pages     = {21-38},
  doi       = {10.1007/978-3-032-06109-6_2},
  url       = {https://mlanthology.org/ecmlpkdd/2025/fieback2025ecmlpkdd-ecd/}
}