ProLaViT: Learning Progressive Latent Visual Thoughts in Structured Latent Space

2026-07-02

AI Summary

This work addresses the challenge that multimodal large language models struggle to jointly coordinate multi-step perception and logical reasoning in complex visual tasks, often relying on external tools or lacking structured cognitive mechanisms. The authors propose ProLaViT, a framework that enables endogenous structured visual reasoning within a continuous latent space. It leverages a self-distillation mechanism, where the model’s own visual encoder supervises its implicit reasoning process, and internalizes algorithmic capabilities through procedurally synthesized data. The approach introduces novel coarse-to-fine causal chains and dialectical reasoning pathways, complemented by a distance-weighted diversity loss that imposes topology-aware constraints to prevent feature collapse. Experiments demonstrate that ProLaViT significantly outperforms existing baselines on core visual reasoning benchmarks, achieving notable advances in accuracy, interpretability, and computational efficiency.

Abstract

Multimodal Large Language Models (MLLMs) have achieved remarkable progress but still struggle with complex visual reasoning tasks requiring multi-step perception and logical deduction. While explicit visual generation incurs prohibitive computational costs, existing latent approaches often rely on external experts or lack rigorous cognitive logic. In this paper, we introduce ProLaViT (Progressive Latent Visual Thought), a framework empowering MLLMs to perform structured visual derivation in the continuous latent space. Unlike works dependent on heterogeneous external models, ProLaViT leverages an endogenous self-distillation mechanism, utilizing the model's own visual encoder to supervise latent thoughts. To facilitate this, we construct a scalable programmatic synthesis pipeline enabling the model to internalize algorithmic precision without inference time tools. We design two reasoning paradigms: (1) Coarse-to-Fine Causal Chain for spatial tasks, guiding attention from global context to local targets. (2) Dialectical Reasoning Chain for logical tasks, incorporating counter-factual thinking for verification. Furthermore, we propose a Distance-Weighted Diversity Loss to impose topology-aware constraints, preventing feature degeneration by enforcing semantic distinctiveness. Extensive experiments demonstrate that ProLaViT outperforms baselines on vision-centric benchmarks, achieving superior accuracy and interpretability with high efficiency.

Problem

• Multimodal Large Language Models
• Visual Reasoning
• Latent Space
• Cognitive Logic
• Computational Cost

Innovation

• Progressive Latent Visual Thought
• Self-Distillation
• Structured Latent Space
• Dialectical Reasoning Chain
• Distance-Weighted Diversity Loss