Learning to Focus and Precise Cropping: A Reinforcement Learning Framework with Information Gaps and Grounding Loss for MLLMs

1. Problem Statement

Multimodal Large Language Models (MLLMs) struggle with fine-grained perception and reasoning in complex, high-resolution visual scenes, particularly when targets are small or obscured. While recent "agent-based" workflows allow MLLMs to use cropping tools to zoom in on regions of interest (ROIs), existing training methods suffer from a critical flaw: formalistic tool invocation.

Empirical analysis reveals that current models (e.g., DeepEyes) often:

• Rely on Global Input: They answer questions based on the full, high-resolution image rather than the cropped region.
• "Answer First, Crop Later": The model predicts the answer before executing the crop, using the crop merely to confirm a pre-existing conclusion rather than to gather new information.
• Lack Precision: Even when they crop, the regions are often imprecise or contain excessive redundant context.

This behavior stems from training data where the global image and the source of the crop share the same resolution, meaning the crop provides no additional information compared to the global view, failing to incentivize the model to actually use the tool.

2. Methodology

The authors propose a two-stage pure Reinforcement Learning (RL) framework that requires no trajectory supervision from a teacher model. The core innovation is forcing the model to rely on cropped details through specific reward mechanisms and data strategies.

Stage 1: Learn to Focus (Information Gap Mechanism)

The goal is to compel the model to actively seek information from cropped regions.

• Offline Resolution Selection: Instead of feeding the full high-resolution image, the system iteratively downsamples the global input image ($I_0$) until the model's answer diverges from the answer derived from the original high-resolution image. This creates an "optimal" low-resolution global view that provides context but lacks fine details.
• Information Gap: The cropped region ($I_{crop}$) is extracted from the original, full-resolution image.
• Mechanism: This creates a crucial "information gap." The low-res global view is insufficient for a correct answer, making the high-detail crop indispensable. The model is rewarded only if it correctly uses the crop to bridge this gap.
• Reward: Includes accuracy, format, and a specific tool-usage bonus granted only when a correct answer is derived using at least one external tool.

Stage 2: Learn to Crop Precisely (Grounding Loss)

The goal is to refine the bounding box precision, as Stage 1 models tend to crop overly large regions to maximize information.

• Hierarchical Annotation: A small set of human-annotated bounding boxes ($B_1, B_2, ... B_n$) is introduced, ranging from the minimal necessary region to a broader context.
• Grounding Reward ($r_{geo}$): A composite reward signal is designed to guide precise cropping:
  1. IoU Reward ($r_{IoU}$): Encourages overlap with Ground Truth (GT) boxes. Crucially, this is conditional: it is only granted if the predicted box covers at least 90% of the minimal GT box ($B_1$). This prevents the model from predicting tiny boxes that miss details.
  2. L1 Distance Reward ($r_{l1}$): Measures the distance between the predicted box corners and the closest GT box. This provides a dense training signal even when the IoU condition is not met, preventing sparse rewards.
• Objective: The model learns to crop the exact necessary area, reducing computational overhead and noise.

3. Key Contributions

1. Diagnosis of "Answer-First" Behavior: The paper identifies and validates that existing RL-based agents often ignore cropped regions, relying instead on the global image due to information redundancy in training data.
2. Information Gap Mechanism: A novel training strategy that artificially creates a dependency on cropped regions by downsampling the global input while keeping the crop high-resolution. This forces the model to learn the utility of the tool without needing expensive teacher trajectories.
3. Two-Stage RL Framework: A trajectory-free approach that first teaches the model to look (Stage 1) and then how to look precisely (Stage 2) using a small amount of annotated data and a composite grounding loss.
4. Efficiency: The method significantly reduces the number of visual tokens required for the global image while maintaining or improving performance.

4. Experimental Results

The method was evaluated on HR-Bench 8K, HR-Bench 4K, and V* benchmarks.

• State-of-the-Art Performance: The proposed method achieved SOTA results on all benchmarks.
  • Under a strict 1,024 visual token budget for the global image, it outperformed models allowed 16,384 tokens (e.g., outperforming Mini-o3 and DeepEyes).
  • It demonstrated superior performance in both Fine-grained Single-instance Perception (FSP) and Cross-instance Perception (FCP).
• Ablation Studies:
  • Stage 1: Improved accuracy by ~8.9% on HR-Bench 8K compared to baselines, proving the Information Gap effectively forces reliance on crops.
  • Stage 2: Further improved accuracy and significantly increased IoU (Intersection over Union) by 5-9%, confirming better localization precision.
  • Data Efficiency: Using a small, challenging dataset (VisualProbe) for Stage 2 was more effective than reusing Stage 1 data.
• Efficiency Analysis:
  • Inference Time: The method was significantly faster (e.g., 2.8s vs 12.4s for DeepEyes on HR-Bench 8K) because it processes the global image at a much lower resolution (1,024 tokens vs 16,384).
  • Training Time: Total training time was lower than DeepEyes due to the reduced token count in the first stage.

5. Significance

This work addresses a fundamental limitation in agentic MLLMs: the tendency to treat tool usage as a formality rather than a necessity. By introducing the Information Gap, the authors demonstrate that models can be trained to genuinely depend on tool-generated observations without expensive teacher-student distillation.

The two-stage approach offers a scalable and efficient path for high-resolution visual reasoning, proving that precise cropping and fine-grained attention can be learned through pure RL with minimal human annotation. This paves the way for more efficient, high-performance MLLMs capable of handling complex real-world visual tasks where details matter.