Why did it reach that result? Establishing Multimodal XAI Technology to Explain Reasoning Grounds

NTT Corporation has established 'Evidence-Enhanced Decoding' technology as a new inference mechanism to improve the reliability of outputs from multimodal AI foundation models that handle images and language. Addressing the issue where LVLMs tend to ignore their own generated reasoning grounds during Chain-of-Thought (CoT) processes, this technology separates and weights inference from images and grounds, unlike conventional methods. This enables the model to output answers by faithfully utilizing information from both sources. This achievement will be presented at the Computer Vision and Pattern Recognition (CVPR) 2026 conference, held in Denver, USA, from June 3 to June 7, 2026. In recent years, while LVLM development has advanced, existing CoT mechanisms left the use of grounds to the model, failing to guarantee consistency between grounds and final outputs. This research establishes a plug-and-play decoding technique that requires no additional training, successfully providing interpretability to the LVLM inference process. This is expected to accelerate social implementation in fields requiring highly reliable systems, such as medical image diagnosis and decision-making support.