Gen-VCoT: Generative Visual Chain-of-Thought Reasoning via Diffusion-Based RGB Intermediate Representations

A new framework called Gen-VCoT proposes generating RGB images as intermediate reasoning steps for multimodal AI models, moving beyond text-only chain-of-thought methods. The approach, detailed in a paper posted to arXiv on June 15, 2026, uses expert vision models to produce visual reasoning traces that researchers say improve performance on spatial and depth-related questions [1][2]. The framework operates in three stages: visual grounding using SAM segmentation, geometric reasoning via Marigold depth maps, and semantic reasoning through integration with the Qwen2-VL model. An adaptive router determines how many reasoning steps are needed for a given query [1][2]. Evaluations showed Gen-VCoT boosted accuracy on spatial questions by 25% and on depth questions by 50% compared to baselines [1][2]. However, the paper also reports a notable limitation: on the CLEVR visual reasoning dataset, standard text-based chain-of-thought achieved 91.2% accuracy while the visual intermediate approach reached only 62.5%, indicating that the optimal reasoning representation depends on the task [1][2]. The authors note the framework may also degrade performance on simple factual queries [2]. The paper was submitted under the Computer Vision and Pattern Recognition category on arXiv, an open-access repository that hosts electronic preprints across physics, mathematics, computer science, and related fields [6]. Founded in 1991, arXiv passed two million articles by the end of 2021 and now receives roughly 24,000 submissions per month as of late 2024 [6]. Papers on the platform are moderated but not peer-reviewed before posting [6]. Multimodal large language models, the class of systems Gen-VCoT targets, are machine learning models with many parameters trained on vast text corpora for language generation and understanding tasks [8]. The Gen-VCoT authors argue that existing multimodal models lack interpretable visual intermediates, relying instead on opaque token representations or external tool calls [2]. By producing human-readable RGB images as reasoning steps, the framework aims to make model decision-making more transparent [1][2]. The paper appears alongside standard arXiv features including Bibliographic Explorer for citation tree navigation and CORE Recommender for discovering related open-access papers [4][5]. These tools are part of arXivLabs, a framework launched in 2020 that allows community collaborators to build experimental features on the platform under guidelines emphasizing openness and user data privacy [4].