TVI-CoT: Text-Visual Interleaved Chain-of-Thought Reasoning for Multimodal Understanding

Researchers have proposed TVI-CoT, a framework for multimodal large language models that enables explicit interleaving of textual reasoning and visual feature access, achieving state-of-the-art results on eight benchmarks.

Chain-of-thought (CoT) reasoning has proven effective for enhancing problem-solving in large language models, but existing CoT approaches suffer from a fundamental limitation: they perform reasoning entirely in text without accessing visual features^[1]. TVI-CoT addresses this limitation by using learnable control tokens , , and to enable dynamic switching between reasoning and visual grounding. Experiments on eight benchmarks demonstrate state-of-the-art results among MLLM-based CoT methods, with notable performance boosts compared to the baseline: +6.1% on MMMU, +3.8% on MathVerse, and +3.4% on both MathVista and ScienceQA^[1]. Meanwhile, a separate study on arxiv.org^[2] found that traditional token-level analysis fails to capture reasoning-level scaling dynamics, and proposed CoT-Space, a novel framework that models the reasoning process as an optimization process in a continuous semantic space. The convergence of CoT length is a natural consequence of the underfitting-overfitting trade-off, according to this study^[2].