[Paper Review] Seeing is Believing: Mitigating Hallucination in Large Vision-Language Models via CLIP-Guided Decoding

Ailin Deng, Z.-H. Chen+1 more|arXiv (Cornell University)|Feb 23, 2024

COVID-19 diagnosis using AIMedicine3 citations

TL;DR

This paper proposes CLIP-Guided Decoding (CGD), a training-free method that reduces object hallucination in Large Vision-Language Models (LVLMs) by using CLIP scores to guide sentence-level decoding. By favoring generated sentences with higher CLIP similarity to the input image, CGD significantly reduces hallucination—achieving a 33.5% relative reduction in hallucination scores—while preserving generation quality and requiring no model fine-tuning or external tools.

ABSTRACT

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.

Motivation & Objective

• To investigate the root causes of object hallucination in LVLMs, particularly in later sentence generations.
• To evaluate whether CLIP similarity serves as a more robust hallucination indicator than internal model likelihoods.
• To develop a training-free, lightweight method that improves hallucination mitigation during decoding using CLIP as external guidance.
• To preserve generation quality while reducing hallucination, especially in safety-critical and human-AI interaction settings.

Proposed method

• The method introduces CLIP-Guided Decoding (CGD), which uses CLIP to score candidate sentences during decoding based on their visual-semantic alignment with the input image.
• At each decoding step, multiple candidate sentences are generated, and their CLIP scores are computed to assess how well they align with the image.
• A hallucination score is computed as a weighted combination of sentence likelihood and CLIP score, with higher CLIP scores prioritizing candidates.
• The final output is selected from the highest-scoring candidates using a differentiable re-ranking mechanism that emphasizes visual grounding.
• The approach is entirely inference-time, requiring no model fine-tuning or additional training.
• The method is evaluated using CLIP models both separate from and identical to those used in the LVLMs, demonstrating robustness across configurations.

Experimental results

Research questions

• RQ1 Is CLIP similarity a stronger and more reliable indicator of hallucination than token likelihood in LVLMs?
• RQ2 Does using CLIP scores to guide decoding reduce hallucination in open-ended image captioning?
• RQ3 Can a training-free, external guidance method like CGD effectively mitigate hallucination without compromising generation quality?
• RQ4 How does the performance of CGD vary with different CLIP models, including those used in the LVLMs themselves?
• RQ5 Does hallucination increase significantly in later sentences of generated captions across different LVLMs?

Key findings

• CLIP scores are a significantly stronger and more robust indicator of hallucination than token likelihoods, especially for later sentences in captions.
• CGD reduces hallucination by 33.5% relative to greedy decoding, achieving a 29.73 hallucination score on COCO and 8.12 on NoCaps.
• The ablation study confirms that CLIP guidance contributes dominantly to hallucination reduction, with removal of CLIP guidance causing performance to degrade to near-greedy levels.
• Even when reusing the same CLIP model used in the LVLM (e.g., OpenAI ViT-L/14), CGD still outperforms baseline decoding, suggesting overfitting in existing fine-tuning.
• Increasing the number of candidates (N) and sampling times (M) improves performance, indicating that broader search with CLIP guidance enhances hallucination mitigation.
• The method preserves generation quality, with average caption length only slightly reduced from 80.05 (greedy) to 76.66 (CGD), indicating minimal degradation in utility.