Constructing causal latent-space reasoning chains

Construct a rigorous, high-quality, causal reasoning chain within the latent space of multimodal large language models.

Background

The authors find that current latent tokens exhibit weak causal influence and limited semantics, suggesting that existing approaches do not yet realize faithful latent-space reasoning.

They explicitly identify as unsolved the problem of how to build causal reasoning chains in latent space, pointing to a challenging direction for future research.