Whether and how MLLMs reason deliberatively in latent space

Establish whether multimodal large language models actually perform deliberative reasoning within their latent space and, if so, characterize how such latent-space reasoning operates.

Background

LVR posits that models deliberate using latent tokens—the last-layer hidden states—during reasoning. However, it is not evident that these hidden states implement a genuine, step-by-step reasoning process.

By framing the process as a causal chain (input → latent tokens → answer), the paper raises the explicit uncertainty of whether deliberate reasoning occurs in latent space and in what manner.