Multimodal Latent Reasoning via Hierarchical Visual Cues Injection
Multimodal Large Language Models (MLLMs) struggle with reasoning because they use fast, often hallucination-prone methods like Chain of Thought. This work introduces HIVE, a framework that enables 'slow thinking' by operating in a latent space. It uses a loop transformer architecture to iteratively refine representations and inject hierarchical visual cues—from global context to fine details—to achieve grounded, multi-step inference.