Multimodal chain-of-thought (MCoT) has become a common way to improve reasoning in multimodal large language models by explicitly modeling intermediate reasoning processes. A new arXiv paper, "From Stages to States: Unified Iterative Reasoning for MLLMs," takes that idea as its starting point and proposes a unified iterative reasoning framework.

According to the abstract, most learning-based MCoT methods have a limitation, although the announcement text cuts off before specifying what it is. The title suggests the fix is to move from treating reasoning as discrete stages to treating it as a continuous state, allowing iteration within a single unified process.

Because the abstract is truncated, the source does not include implementation details, benchmark results, or comparisons with existing MCoT variants. The contribution is therefore best understood at the level of framing: a call to unify iterative reasoning for multimodal models.