Zero-Shot Semantic Segmentation Research of Vision Language Models

In this paper, we systematically research recent based on vision language models (VLMs) semantic segmentation methods, focusing on two major paradigms: VLM-based crossmodal models and large language model (LLM)-enhanced interactive models. We analyze the characteristics and strategies of representative methods in these two paradigms, covering zero-shot learning, visual-language prompting strategies, and multi-round interactive reasoning. We summarize and analyze their segmentation accuracy and computational performance, and show that VLM-based crossmodal models remain competitive in structured datasets due to their efficiency and simplicity, while LLM-enhanced methods show greater flexibility and reasoning capabilities in complex instruction-driven tasks. Our study highlights the advantages of both paradigms and proposes a future direction of combining lightweight visual foundations with high-level semantic reasoning.