Multimodal large language models (MLLMs) typically reason in a "fast thinking" mode, either generating answers end-to-end or relying on explicit, language-centric chains of thought. The authors of the new paper argue that this approach is inefficient, verbose, and prone to hallucination. They propose HIVE (multimodal latent reasoning via HIerarchical Visual cuEs injection), a framework that performs deliberate, multi-step inference entirely in an aligned latent space, without producing textual rationales.

HIVE works by recursively extending transformer blocks, creating an internal loop that refines reasoning over multiple iterations. Crucially, the model grounds this process by injecting hierarchical visual cues—from global scene context down to fine-grained regional details—directly into its latent representations. This grounding allows the model to incorporate vision knowledge at each step of its internal reasoning.

In evaluations, the authors report that test-time scaling is effective when vision knowledge is included, and that integrating hierarchical information significantly improves performance on complex scene understanding. The work positions latent-space reasoning as a promising alternative to explicit chain-of-thought methods for multimodal tasks, though the paper's abstract does not include quantitative comparisons or benchmark details.