Visual-Text Dual Chain-of-Thought

• Visual-Text Dual Chain-of-Thought is a multimodal reasoning framework that combines visual evidence with textual inference to improve decision-making.
• It employs diverse strategies such as interleaved image-text sequences, dynamic visual insertion, and parallel visual-linguistic streams to maintain continuous visual grounding.
• The approach advances reasoning accuracy by integrating intermediate visual representations with explicit text-based analysis during each inference step.

Searching arXiv for the specified paper and closely related work on visual-text dual chain-of-thought reasoning. Visual-Text Dual Chain-of-Thought denotes a class of multimodal reasoning frameworks in which intermediate inference is not confined to a purely textual rationale, but instead couples textual reasoning with visual evidence, visual states, or visual intermediate representations. In this literature, the duality can take several forms: interleaving textual steps with image regions or frames, maintaining parallel visual and linguistic reasoning streams, rendering reasoning into a visual canvas, or replacing explicit word sequences with latent visual-textual states. A recurring motivation is that text-only multimodal reasoning often becomes “vision-blind reasoning,” whereas dual CoT seeks to preserve visual grounding throughout the reasoning process rather than only at initial encoding (Hu et al., 7 Jun 2026). A related unifying claim is that multimodal CoT improves large vision-LLMs by introducing “visual thoughts,” namely intermediate representations that convey image information into the reasoning process, regardless of whether the outward format is textual or interleaved image-text output (Cheng et al., 21 May 2025).

1. Historical emergence and conceptual scope

Early work established the premise that CoT in vision-language settings should not be treated as a trivial transfer of language-only prompting. “Chain of Thought Prompt Tuning in Vision LLMs” introduced a chain of prompts and corresponding networks, with a dynamic chain controller, and stated that it was the first work to successfully adapt chain-of-thought prompting that combines visual and textual embeddings (Ge et al., 2023). In parallel, “Visual Chain of Thought” proposed recursively generating multimodal infillings to bridge logical gaps in sequential data, making explicit the idea that intermediate steps could be both textual and visual rather than textual alone (Rose et al., 2023).

Subsequent work expanded this premise from prompt design to explicit multimodal reasoning substrates. “Chain of Images for Intuitively Reasoning” argued that current models merely align images into textual space and proposed using a series of generated images as intermediate representations; it also introduced the CoI evaluation dataset across 15 distinct domains where images can intuitively aid problem-solving (Meng et al., 2023). “Cantor” reframed multimodal CoT as a perception-decision architecture in which visual context acquisition and logical reasoning are jointly organized through expert-role prompting inside a single MLLM, without fine-tuning or ground-truth rationales (Gao et al., 2024).

By 2025–2026, the field had diversified into several distinct but related families: hierarchical grounding for 3D vision-language learning (Chen et al., 8 Mar 2025), interleaved text-visual reasoning with explicit control tokens (Hu et al., 7 Jun 2026), dynamic visual-thought insertion (Liu et al., 23 Mar 2026), mutable visual state reasoning (Sun et al., 11 Feb 2026), latent-space or continuous-thought approaches (Shao et al., 31 Jan 2026, Pham et al., 18 Aug 2025), and parallel visual-linguistic reasoning for action-conditioned systems (Zhong et al., 23 Mar 2026). Across these lines, “dual CoT” no longer refers to a single architecture, but to the broader principle that visual information must remain available, revisable, and structurally coupled to reasoning.

2. Representational forms and architectural families

A useful conceptual synthesis is provided by “Visual Thoughts: A Unified Perspective of Understanding Multimodal Chain-of-Thought,” which distinguishes Textual-MCoT, taking multimodal input and producing textual output, from Interleaved-MCoT, which generates interleaved image-text outputs (Cheng et al., 21 May 2025). The same work defines four forms of visual thought expression: Natural Language, Structured Language, Edited Image, and Generative Image. Its central claim is that these forms differ in clarity and conciseness, and that MCoT gains depend on those properties rather than on modality labels alone (Cheng et al., 21 May 2025). This suggests that dual CoT is best understood as a design space of intermediate representations rather than a single prompting convention.

Several architectural families instantiate that design space:

Paradigm	Representative work	Defining mechanism
Textual visual thoughts	“Visual Thoughts” (Cheng et al., 21 May 2025)	Natural Language or Structured Language as intermediate visual thought expressions
Explicit interleaving	“TVI-CoT” (Hu et al., 7 Jun 2026)	Learnable control tokens <THINK>, <LOOK>, <ANSWER>
Dynamic visual insertion	“DaP-ICoT” (Liu et al., 23 Mar 2026)	Confidence-triggered visual thought integration plus precise visual guidance
Mutable external state	“Canvas-of-Thought” (Sun et al., 11 Feb 2026)	HTML Canvas with DOM-based CRUD operations
Latent modal mixing	“Learning Modal-Mixed Chain-of-Thought Reasoning with Latent Embeddings” (Shao et al., 31 Jan 2026)	Text tokens interleaved with compact visual sketches represented as latent embeddings
Continuous latent reasoning	“MCOUT” (Pham et al., 18 Aug 2025)	Iteratively refined continuous hidden vector in joint latent space
Parallel dual-branch reasoning	“DualCoT-VLA” (Zhong et al., 23 Mar 2026)	Visual CoT and linguistic CoT via separate learnable query tokens

The interleaving family makes the visual-text alternation explicit. TVI-CoT augments a base MLLM with control tokens and a visual grounding module, so that the model alternates between reasoning and visual access during generation. At each <LOOK> step, reasoning-conditioned attention over visual tokens is computed as

$$\alpha_t = \text{softmax}\left(\frac{\mathbf{h}_t \mathbf{W}_q (\mathbf{v} \mathbf{W}_k)^\top}{\sqrt{d}}\right),$$

and the top-$k$ attended visual tokens with $k=32$ are retrieved as the grounding set (Hu et al., 7 Jun 2026).

Other families treat dual Co