Unclear mechanisms behind latent visual reasoning effectiveness

Determine the underlying mechanisms that drive the effectiveness of latent visual reasoning in multimodal large language models.

Background

Latent Visual Reasoning (LVR) asks multimodal LLMs to reason via hidden-state "latent tokens" aligned with visual embeddings. Although recent methods report promising empirical results, the specific internal dynamics by which these latent tokens contribute to performance are not well understood.

The paper employs causal mediation analysis to probe the causal chain from inputs to latent tokens to outputs, motivating the explicit question of what mechanisms, if any, underlie LVR’s reported effectiveness.