CHASD: Language Increment-Calibrated Contrastive Decoding against Hallucination in LVLMs
The paper presents CHASD, a new framework designed to reduce hallucinations in Large Vision-Language Models (LVLMs). It introduces a method that activates contrastive decoding only when necessary, improving inference efficiency. Experimental results demonstrate that CHASD outperforms existing training-free baselines in hallucination-related metrics.
- ▪CHASD stands for Contrastive Hallucination-Aware Step-wise Decoding.
- ▪The framework uses an uncertainty-driven confidence gate to optimize decoding steps.
- ▪Experiments show that CHASD improves performance on several benchmarks.
Opening excerpt (first ~120 words) tap to expand
Computer Science > Computer Vision and Pattern Recognition arXiv:2605.23344 (cs) [Submitted on 22 May 2026] Title:CHASD: Language Increment-Calibrated Contrastive Decoding against Hallucination in LVLMs Authors:Xiaoyi Huang, Kejia Zhang, Zhiming Luo View a PDF of the paper titled CHASD: Language Increment-Calibrated Contrastive Decoding against Hallucination in LVLMs, by Xiaoyi Huang and 2 other authors View PDF HTML (experimental) Abstract:Large Vision-Language Models have shown strong multimodal reasoning capabilities, yet they remain susceptible to object hallucinations when language priors dominate insufficient or misaligned visual evidence.
…
Excerpt limited to ~120 words for fair-use compliance. The full article is at arXiv cs.AI.
Discussion
0 commentsMore from arXiv cs.AI
