ECD: Efficient Contrastive Decoding with Probabilistic Hallucination Detection

doi:10.1007/978-3-032-06109-6_2

ECD: Efficient Contrastive Decoding with Probabilistic Hallucination Detection

2026-01-01

0

PRE

AI

DOI：10.1007/978-3-032-06109-6_2

Original

Original request for help

Save

Abstract

En 中文

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.

Keywords:

Multimodal large language models

Contrastive decoding

Hallucination mitigation

Hallucination detection

Meta classification

ECD: Efficient Contrastive Decoding with Probabilistic Hallucination Detection

Abstract

Journal

Organization