SplitGround: Long-Chain Reasoning Split via Modular Multi-Expert Collaboration for Training-Free Scene Knowledge-Guided Visual Grounding

Scene Knowledge-guided Visual Grounding (SK-VG) is a multi-modal detection task built upon conventional visual grounding (VG) for human–computer interaction scenarios. It utilizes an additional passage of scene knowledge apart from the image and context-dependent textual query for referred object localization. Due to the inherent difficulty in directly establishing correlations between the given query and the image without leveraging scene knowledge, this task imposes significant demands on a multi-step knowledge reasoning process to achieve accurate grounding. Off-the-shelf VG models underperform under such a setting due to the requirement of detailed description in the query and a lack of knowledge inference based on implicit narratives of the visual scene. Recent Vision–Language Models (VLMs) exhibit improved cross-modal reasoning capabilities. However, their monolithic architectures, particularly in lightweight implementations, struggle to maintain coherent reasoning chains across sequential logical deductions, leading to error accumulation in knowledge integration and object localization. To address the above-mentioned challenges, we propose SplitGround—a collaborative framework that strategically decomposes complex reasoning processes by fusing the input query and image with knowledge through two auxiliary modules. Specifically, it implements an Agentic Annotation Workflow (AAW) for explicit image annotation and a Synonymous Conversion Mechanism (SCM) for semantic query transformation. This hierarchical decomposition enables VLMs to focus on essential reasoning steps while offloading auxiliary cognitive tasks to specialized modules, effectively splitting long reasoning chains into manageable subtasks with reduced complexity. Comprehensive evaluations on the SK-VG benchmark demonstrate the significant advancements of our method. Remarkably, SplitGround attains an accuracy improvement of 15.71% on the hard split of the test set over the previous training-required SOTA, using only a compact VLM backbone without fine-tuning, which provides new insights for knowledge-intensive visual grounding tasks.