Enhancing Caption Fidelity via Explanation-Guided Captioning with Vision-Language Fine-Tuning

Image captioning models have achieved remarkable progress with the introduction of attention mechanisms and transformer-based architectures. However, understanding and diagnosing their predictions remain a challenging task, particularly in terms of attribution, interpretability, and mitigation of hallucinated outputs. In this work, we present \textbf{CAPEV}, a novel explanation-guided fine-tuning paradigm that builds upon Layer-wise Relevance Propagation (LRP) to improve caption reliability and semantic grounding. We begin by systematically adapting state-of-the-art explanation methods—including LRP, Grad-CAM, and Guided Grad-CAM—to image captioning architectures with both adaptive and multi-head attention mechanisms. Unlike conventional attention heatmaps, which offer a coarse visual explanation, these gradient-based and propagation-based methods provide dual-perspective relevance: spatial pixel-level attributions for image regions and token-wise linguistic relevance across sequential inputs. Through rigorous comparisons, we find that these methods yield a more precise and disentangled understanding of the model's decision basis. Building on these insights, we introduce CAPEV, an inference-time fine-tuning approach that leverages explanation signals to recalibrate the internal representations of the model. By identifying both supporting and opposing relevance cues for each word prediction, CAPEV dynamically adjusts context features to suppress hallucinated entities and reinforce grounded content. Notably, CAPEV operates without requiring additional external annotations or human supervision. Extensive experiments on Flickr30K and MSCOCO benchmarks demonstrate that CAPEV significantly reduces object hallucination while preserving caption fluency and overall performance on standard evaluation metrics. Our findings suggest that integrating explainability into the training loop opens a promising avenue toward transparent and trustworthy vision-language generation.