Simple-VGC: Enhancing Visual Grounding in Multimodal Reasoning via Adaptive Tool Composition

Multimodal large language models (MLLMs) have achieved strong performance on visionlanguage tasks, yet often fail to preserve and effectively leverage visual evidence throughout generation. We identify three fundamental types of visual grounding failures: Long-Context Grounding Error, where visual information gradually decays over long sequences; Fine-Grained Grounding Error, where lowresolution or degraded inputs hinder the recovery of detailed visual information; and Regional Grounding Error, where spatially diffuse attention weakens region-level vision-language alignment. To address these issues, we propose a tool-augmented reasoning framework with three targeted compensation strategies: reuse, which re-injects the original image to mitigate visual forgetting; focus_area, which constrains attention to task-relevant regions; and zoom_in, which enhances visual resolution for fine-grained perception. We further construct the TWI-Tools-146K dataset and develop SIMPLE-VGC, a tool-augmented MLLM that interleaves visual and textual tokens. Extensive experiments show that each tool yields targeted improvements for its corresponding grounding error, while their combination produces synergistic gains in visual reasoning. Beyond performance, our analysis provides mechanistic insights into how tool-based interventions improve visual grounding, pointing toward more reliable multimodal reasoning. Query: Identify the city-zone type, infer the traffic control state, and predict the driver's next action given the current scene. Response: The scene shows a dense urban corridor ....... While the architecture confirms a central business district, the model infers that the broader route transitions to a highway-like stretch immediately ahead ❌ , and thus the driver will accelerate through an already-green light ❌ to maintain flow. Query: How many vehicles are visible in the image, including all lanes? Response: I can see around three vehicles ❌ on the road -one van in the middle and two cars ahead of it ❌ . The rest of the area is somewhat blurred, and I can't clearly tell if there are more vehicles further away ❌ . Long-Context Grounding Error Fine-Grained Grounding Error Regional Grounding Error 1448 x 940 724 x 470 Query: Analyze the relative spatial positions of the vehicles and reason which lane the "NATIONAL HME" van is occupying. Explain whether the white pickup truck is ahead of or behind it, and how their relative positions might affect the van's next maneuver. Response: The "NATIONAL HME" van appears to be in the right lane ❌ , positioned behind ❌ the white pickup truck. Since the pickup is slightly ahead ❌ , it will likely move first, and the van may merge left ❌ afterward to continue along the main road.