Visual Enhanced Depth Scaling for Multimodal Latent Reasoning

TL;DR

This paper introduces a novel approach to multimodal latent reasoning that enhances visual perception and reasoning depth through a visual replay module and routing depth scaling, achieving state-of-the-art results.

Contribution

It proposes a visual replay module and adaptive depth scaling to improve visual token optimization and complex reasoning in multimodal latent models.

Findings

Achieves state-of-the-art performance on multiple benchmarks.

Provides significant inference speedups over explicit CoT methods.

Addresses visual under-optimization and token complexity issues.

Abstract

Multimodal latent reasoning has emerged as a promising paradigm that replaces explicit Chain-of-Thought (CoT) decoding with implicit feature propagation, simultaneously enhancing representation informativeness and reducing inference latency. By analyzing token-level gradient dynamics during latent training, we reveal two critical observations: (1) visual tokens exhibit significantly smaller gradient norms than their textual counterparts due to inherent language bias, resulting in systematic visual under-optimization; and (2) semantically simple tokens converge rapidly, whereas complex tokens exhibit persistent gradient instability constrained by fixed architectural depths. To address these limitations, we propose a visual replay module and routing depth scaling to collaboratively enhance visual perception and refine complicated latents for deeper contextual reasoning. The former module leverages causal self-attention to estimate token saliency, reinforcing fine-grained grounding through spatially-coherent constraints. Complementarily, the latter mechanism adaptively allocates additional reasoning steps to complex tokens, enabling deeper contextual refinement. Guided by a curriculum strategy that progressively internalizes explicit CoT into compact latent representations, our framework achieves state-of-the-art performance across diverse benchmarks while delivering substantial inference speedups over explicit CoT baselines.