Multimodal CoT Reasoning

Multimodal chain-of-thought reasoning asks how models can combine images, video, audio, and text across intermediate steps to answer questions, solve problems, or act. Common approaches include supervised fine-tuning and distillation on reasoning traces, reinforcement learning with step-level or outcome rewards, and interleaving text with visual evidence, tool use, or compact latent states. Benchmarks test settings such as visual mathematics, spatial reasoning, video understanding, charts, documents, and medical imaging, increasingly evaluating whether each step is grounded in evidence rather than only whether the final answer is correct. A central tension is that explicit reasoning can improve complex problem solving but may also increase cost, rely on textual shortcuts, or weaken visual grounding, so adaptive and more faithful reasoning remains an open goal.