Ailin Deng

Ailin Deng, Zhirui Chen, Bryan Hooi

Large Vision-Language Models (LVLMs) are susceptible to object hallucinations, an issue in which their generated text contains non-existent objects, greatly limiting their reliability and practicality. Current approaches often rely on the model's token likelihoods or other internal information, instruction tuning on additional datasets, or incorporating complex external tools. We first perform empirical analysis on sentence-level LVLM hallucination, finding that CLIP similarity to the image acts as a stronger and more robust indicator of hallucination compared to token likelihoods. Motivated by this, we introduce our CLIP-Guided Decoding (CGD) approach, a straightforward but effective training-free approach to reduce object hallucination at decoding time. CGD uses CLIP to guide the model's decoding process by enhancing visual grounding of generated text with the image. Experiments demonstrate that CGD effectively mitigates object hallucination across multiple LVLM families while preserving the utility of text generation. Codes are available at https://github.com/d-ailin/CLIP-Guided-Decoding.

Xixi Liu, Ailin Deng, Christopher Zach

Mitigating object hallucination in large vision-language models (LVLMs) is critical to their safe deployment. Existing methods either are restricted to specific decoding methods, or demand sophisticated modifications to visual inputs, or rely on knowledge from external models. In this work, we first reveal the phenomenon that VLMs exhibit significant imbalance in the ``Yes'' ratio ( \ie, the fraction of ``Yes'' answers among the total number of questions) across three different visual question answering (VQA) datasets. Furthermore, we propose an energy-based decoding method, which dynamically selects the hidden states from the layer with minimal energy score. It is simple yet effective in reducing the bias for the yes ratio while boosting performance across three benchmarks (POPE, MME, and MMVP). Our method consistently improves accuracy and F1 score on three VQA datasets across three commonly used VLMs over several baseline methods. The average accuracy improvement is 4.82% compared to greedy decoding. Moreover, the average yes-ratio gap reduction is 8.81%, meaning the proposed method is less biased as shown in Figure 1.