ProLaViT enables MLLMs to perform structured visual derivation in latent space

The paper introduces ProLaViT (Progressive Latent Visual Thought), a framework that empowers Multimodal Large Language Models to perform structured visual derivation within a continuous latent space. Unlike approaches relying on external experts, it uses an endogenous self-distillation mechanism where the model's own visual encoder supervises latent thoughts.

• It utilizes a scalable programmatic synthesis pipeline to internalize algorithmic precision without inference-time tools.
• The framework designs two reasoning paradigms: Coarse-to-Fine Causal Chain for spatial tasks and Dialectical Reasoning Chain for logical tasks.
• A Distance-Weighted Diversity Loss is proposed to impose topology-aware constraints and prevent feature degeneration.

Experiments demonstrate that ProLaViT outperforms baselines on vision-centric benchmarks, achieving superior accuracy and interpretability with high efficiency.