Research map / Vision-language and multimodal / Multimodal LLMs

ActiveScope: Actively Seeking and Correcting Perception for MLLMs

ICML 2026 · Vision-language and multimodal › Multimodal LLMs · approach: Instruction-tuned multimodal assistants · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes ActiveScope, a training-free framework that localizes multiple query targets via keyword-anchored attention (SAL) and iteratively suppresses attention on verified-wrong regions (ISR) before zoomed-in answering on high-resolution benchmarks.

Problem

MLLMs struggle with fine-grained perception in high-resolution images; attention-based localization is misled by salient distractors (contextual dominance) and fails to locate all targets in multi-object queries (semantic bias).

Approach

Use per-target keyword token attention (normalized by a generic-instruction reference attention) as semantic anchors to localize each object independently, then verify crops with the MLLM and mask attention to failed regions via a hard attention mask in an iterative loop before composing a zoomed view.

Key results

• On V* Bench, ActiveScope with Qwen3-VL 4B reaches 96.34% average accuracy, vs. 89.53% regular and 93.19% for ZoomEye.
• With InternVL2.5 4B, ActiveScope reaches 80.63% on V* Bench (vs. 68.59% regular) and 67.63% on HR-Bench 8K (vs. 64.50% ZoomEye).
• Efficiency: ActiveScope averages 4.7 VQA calls on V* Bench vs. 33 for ZoomEye, with 24 min vs. 27 min total inference time and ~4GB more GPU memory.

Surprising findings

• Iterative localization with Top-K re-localization or interference suppression yields only modest gains over the baseline, while key-token guidance gives a much larger localization F1 jump (40.9 to 65.0).

Limitations

• Relies on rectangular bounding boxes, which are suboptimal for irregularly shaped targets.
• Requires access to internal attention maps, so the closed-source adaptation is only approximate and gains are small (GPT-4o 65.97 to 66.75).
• May not generalize to every image category or specialized downstream task.

Experimental setup

Datasets: V* Bench HR-Bench 4K HR-Bench 8K MME-RealWorld-Lite

Benchmarks: V* Bench HR-Bench 4K HR-Bench 8K MME-RealWorld-Lite TextVQA DocVQA GQA VSR

Metrics: accuracy F1 (IoGT) precision recall

Baselines: ViCrop ICoT ZoomEye Qwen-VL-max GPT-4o

Builds on: ViCrop ZoomEye relative attention

Closest papers by meaning

• Mechanisms of Object Localization in Vision-Language Models (CVPR 2026 · similarity 0.88)
• HiDe: Rethinking The Zoom-IN method in High Resolution MLLMs via Hierarchical Decoupling (ICML 2026 · similarity 0.88)
• BVS: Bayesian Visual Search with Multimodal Large Language Model for Fine-grained Perception (ICML 2026 · similarity 0.88)
• Towards Long-window Anchoring in Vision-Language Model Distillation (AAAI 2026 · similarity 0.88)
• Catching the Details: Self-Distilled RoI Predictors for Fine-Grained MLLM Perception (ICLR 2026 · similarity 0.87)
• Your Large Vision-Language Model Only Needs A Few Attention Heads For Visual Grounding (CVPR 2025 · similarity 0.87)