Vision-aligned Latent Reasoning for Multi-modal Large Language Model

TL;DR

This paper introduces VaLR, a reasoning framework that aligns visual information in latent space to improve multi-modal large language models' performance on complex, long-context tasks.

Contribution

VaLR dynamically generates vision-aligned latent tokens to enhance reasoning, addressing visual information dilution in existing multi-modal models.

Findings

VaLR outperforms existing methods on long-context understanding benchmarks.

Performance on VSI-Bench improves from 33.0% to 52.9%.

Achieves a 19.9% gain over Qwen2.5-VL.

Abstract

Despite recent advancements in Multi-modal Large Language Models (MLLMs) on diverse understanding tasks, these models struggle to solve problems which require extensive multi-step reasoning. This is primarily due to the progressive dilution of visual information during long-context generation, which hinders their ability to fully exploit test-time scaling. To address this issue, we introduce Vision-aligned Latent Reasoning (VaLR), a simple, yet effective reasoning framework that dynamically generates vision-aligned latent tokens before each Chain of Thought reasoning step, guiding the model to reason based on perceptual cues in the latent space. Specifically, VaLR is trained to preserve visual knowledge during reasoning by aligning intermediate embeddings of MLLM with those from vision encoders. Empirical results demonstrate that VaLR consistently outperforms existing approaches across a wide range of benchmarks requiring long-context understanding or precise visual perception, while exhibiting test-time scaling behavior not observed in prior MLLMs. In particular, VaLR improves the performance significantly from 33.0% to 52.9% on VSI-Bench, achieving a 19.9%p gain over Qwen2.5-VL.