CrystaL: Spontaneous Emergence of Visual Latents in MLLMs

TL;DR

CrystaL is a novel framework that enhances visual reasoning in multimodal large language models by explicitly aligning latent representations for better visual understanding without extra annotations.

Contribution

It introduces a single-stage latent alignment method that crystallizes visual semantics in MLLMs, improving fine-grained visual reasoning without auxiliary supervision.

Findings

Outperforms state-of-the-art baselines on perception benchmarks.

Achieves significant improvements in visual understanding tasks.

Maintains robust reasoning capabilities across tests.

Abstract

Multimodal Large Language Models (MLLMs) have achieved remarkable performance by integrating powerful language backbones with large-scale visual encoders. Among these, latent Chain-of-Thought (CoT) methods enable implicit reasoning in continuous hidden states, facilitating seamless vision-language integration and faster inference. However, existing heuristically predefined supervision signals in latent CoT provide limited guidance for preserving critical visual information in intermediate latent states. To address this limitation, we propose CrystaL (Crystallized Latent Reasoning), a single-stage framework with two paths to process intact and corrupted images, respectively. By explicitly aligning the attention patterns and prediction distributions across the two paths, CrystaL crystallizes latent representations into task-relevant visual semantics, without relying on auxiliary annotations or external modules. Extensive experiments on perception-intensive benchmarks demonstrate that CrystaL consistently outperforms state-of-the-art baselines, achieving substantial gains in fine-grained visual understanding while maintaining robust reasoning capabilities.