Reproduce “ActiveScope: Actively Seeking and Correcting Perception for MLLMs” on OpenResearch

About this paper

Yajing Wang, Chao Bi, Junshu Sun, Shufan Shen, Zhaobo Qi, Shuhui Wang, Qingming Huang

ActiveScope, a training-free framework developed by researchers at the University of Chinese Academy of Sciences and Harbin Institute of Technology, enhances Multimodal Large Language Models' (MLLMs) fine-grained perception in high-resolution images by systematically addressing localization failures from contextual dominance and semantic bias. It achieved consistent performance improvements across various MLLMs and benchmarks, boosting Qwen3-VL 4B's accuracy on V*Bench by 6.81% and demonstrating superior efficiency compared to existing search-based methods.

Key insights

• Systematic diagnosis revealed that "contextual dominance" (salient distractors) is the primary bottleneck for inaccurate localization, while "semantic bias" (global query semantics) hinders complete multi-object grounding.
• Leveraging fine-grained semantic keywords as anchors to derive object-specific attention maps can effectively decouple and independently localize multiple targets, overcoming semantic bias.
• An active, self-correcting paradigm involving verification and explicit suppression of attention to misleading regions significantly refines localization precision and robustness.

Summary by alphaXiv.