Learning Modal-Mixed Chain-of-Thought Reasoning with Latent Embeddings

TL;DR

This paper introduces modal-mixed CoT, a multimodal reasoning method that interleaves text with visual sketches using latent embeddings, improving performance on vision-intensive tasks.

Contribution

It presents a novel approach combining language and visual latent embeddings with diffusion decoding, enhancing multimodal reasoning capabilities of LLMs and VLMs.

Findings

Outperforms language-only and other CoT methods on 11 tasks

Effective semantic alignment of visual latent space with VLM

Disentangles perceptual details from high-level reasoning

Abstract

We study how to extend chain-of-thought (CoT) beyond language to better handle multimodal reasoning. While CoT helps LLMs and VLMs articulate intermediate steps, its text-only form often fails on vision-intensive problems where key intermediate states are inherently visual. We introduce modal-mixed CoT, which interleaves textual tokens with compact visual sketches represented as latent embeddings. To bridge the modality gap without eroding the original knowledge and capability of the VLM, we use the VLM itself as an encoder and train the language backbone to reconstruct its own intermediate vision embeddings, to guarantee the semantic alignment of the visual latent space. We further attach a diffusion-based latent decoder, invoked by a special control token and conditioned on hidden states from the VLM. In this way, the diffusion head carries fine-grained perceptual details while the VLM specifies high-level intent, which cleanly disentangles roles and reduces the optimization pressure of the VLM. Training proceeds in two stages: supervised fine-tuning on traces that interleave text and latents with a joint next-token and latent-reconstruction objective, followed by reinforcement learning that teaches when to switch modalities and how to compose long reasoning chains. Extensive experiments across 11 diverse multimodal reasoning tasks, demonstrate that our method yields better performance than language-only and other CoT methods. Our code will be publicly released.