Thinking with Comics: Enhancing Multimodal Reasoning through Structured Visual Storytelling

TL;DR

Thinking with Comics (TwC) addresses the efficiency gap in multimodal reasoning by using comics as a high-density intermediate representation that preserves temporal dynamics while avoiding the redundancy of video. The authors present two pathways—Path I (end-to-end visualized reasoning) and Path II (comics as conditioning context for a VLM)—and show that TwC yields systematic gains on multimodal reasoning and long-context understanding benchmarks, with substantial efficiency advantages over video-based approaches. Key findings include the impact of narrative style and embedded text on performance, a panel-count scaling behavior with optimal 4–6 panels, the importance of preserving temporal order, and strong cross-model generalization across diverse architectures. Collectively, TwC demonstrates a practical, model-agnostic approach to efficient temporal reasoning, with promising implications for long-context reasoning and controllable visual cognition.

Abstract

Chain-of-Thought reasoning has driven large language models to extend from thinking with text to thinking with images and videos. However, different modalities still have clear limitations: static images struggle to represent temporal structure, while videos introduce substantial redundancy and computational cost. In this work, we propose Thinking with Comics, a visual reasoning paradigm that uses comics as a high information-density medium positioned between images and videos. Comics preserve temporal structure, embedded text, and narrative coherence while requiring significantly lower reasoning cost. We systematically study two reasoning paths based on comics and evaluate them on a range of reasoning tasks and long-context understanding tasks. Experimental results show that Thinking with Comics outperforms Thinking with Images on multi-step temporal and causal reasoning tasks, while remaining substantially more efficient than Thinking with Video. Further analysis indicates that different comic narrative structures and styles consistently affect performance across tasks, suggesting that comics serve as an effective intermediate visual representation for improving multimodal reasoning.

Figures (9)

• Figure 1: The selected reasoning tasks and (Long) Context Understanding tasks, along with the Thinking with Comics solution based on Gemini-3 Pro Image. The reasoning tasks primarily involve mathematical and logical reasoning, while the (Long) Context Understanding tasks require the model to comprehend cultural contexts, documents, and other extended information. The model provides the reasoning process and correct answers within the generated comic panels.
• Figure 2: Overview of the two paths of Thinking with Comics paradigm. Path 1 directly utilizes an image generation model to create a comic, where the process of generating the comic constitutes the reasoning process for the problem, and the answer is obtained by extracting the final panel of the comic. Path 2 takes the generated comic along with the original problem as context and inputs them into a VLM, which then performs reasoning and outputs the answer.
• Figure 3: The performance-cost curve across different panel counts $N$. Accuracy enters a plateau at $N \in [4, 6]$. On the MATH500 dataset, token cost ranges between 1100 and 1300.
• Figure 4: Frequency distribution of generated panels across tasks with varying difficulty levels. The shift to the right indicates the model's adaptive allocation of reasoning steps for complex tasks.
• Figure 5: Effect of temporal perturbations on comic-based reasoning. Accuracy under Complete Shuffle (blue) and Intermediate Deletion (orange) decreases as perturbation intensity increases, with deletion causing a larger drop than shuffling.