A recent study shows that most multimodal large language models (MLLMs) feed a single visual summary into the language backbone, ignoring that different language layers need different image details. The authors introduce Dynamic Gated Pathways (DGP), a mechanism that lets every layer of the language model choose which layer of a vision transformer should contribute, based on its current semantic state. When built on the LLaVA‑1.5 7B and 13B systems, DGP improves benchmark scores by +4.2 on SQA‑I, +3.3 on MM‑Vet, +3.1 on VizWiz, and +3.1 on LLaVA‑W compared with the original LLaVA‑1.5 model. These gains come with a small hardware cost: for the 7B model, DGP adds 5.3% more parameters, 4.1% more floating‑point operations, and reduces peak memory by 0.13%. The 13B version sees a 4.3% increase in parameters, negligible extra computation, and a 4.1% memory rise. The authors note that the dynamic routing also gives insight into which image scales each language layer prefers.

The experiments were run at a batch size of one during end‑to‑end autoregressive generation, ensuring that the reported FLOPs and memory figures reflect real‑time inference.