FuguReport
Daily Report / 2026-07-01

Daily Paper Briefing

2026-07-01 1 Paper

Perceive-to-Reason: Decoupling Perception and Reasoning for Fine-Grained Visual Reasoning

Authors Hongxing Li, Xiufeng Huang, Dingming Li, Wenjing Jiang, Zixuan Wang, Haolei Xu, Hanrong Zhang, Haiwen Hong, Longtao Huang, Hui Xue, Weiming Lu, Jun Xiao, Yueting Zhuang, Yongliang Shen
Affiliations Zhejiang University / Alibaba
Categories Method / Visual Reasoning / Decoupling perception and reasoning, Method / Reinforcement Learning / Role-aware update strategy, Application / Vision and Language / Question answering on annotated images
License CC BY 4.0

Abstract Overview

This paper presents Perceive-to-Reason (P2R), a framework for fine-grained visual reasoning that explicitly separates perception from reasoning. In P2R, the model first localizes question-relevant evidence as a Perceiver and then answers the question as a Reasoner using the annotated image and cropped regions. To train this decoupled pipeline under answer-only supervision, the authors introduce PRA-GRPO, a role-aware reinforcement learning strategy that alternates between perception-focused and reasoning-focused updates while keeping the other role fixed. The method is built on Qwen3-VL-Instruct backbones and evaluated on high-resolution visual reasoning benchmarks as well as a broader multimodal benchmark.

Novelty

The main novelty is the explicit two-stage formulation of fine-grained visual reasoning as separate evidence localization and answer generation within a unified model. The paper also introduces PRA-GRPO, an alternating role-aware RL strategy that uses only final-answer supervision to train both stages without ground-truth bounding boxes.

Results

Across Qwen3-VL-Instruct 2B, 4B, and 8B backbones, P2R consistently improves performance on V-Star, HR-Bench-4K, and HR-Bench-8K. For example, P2R-4B reaches 93.2% on V-Star, 81.9% on HR-Bench-4K, and 80.5% on HR-Bench-8K, and the paper reports average gains over the corresponding backbones of 8.1%, 11.0%, and 9.7% for the 2B, 4B, and 8B variants, respectively. The gains also extend to MME-RealWorld-Lite and a ReasonSeg grounding evaluation, indicating broader transfer beyond the primary high-resolution setting.

Key Points

  1. P2R reformulates fine-grained visual reasoning as a two-stage process in which a Perceiver localizes relevant evidence before a Reasoner answers the question from annotated and cropped visual inputs.
  2. PRA-GRPO alternates perception and reasoning updates under shared parameters, converting final-answer correctness into a more attributable training signal without requiring bounding box annotations.
  3. Empirically, the method improves multiple Qwen3-VL-Instruct scales on high-resolution benchmarks and also shows gains on broader multimodal reasoning and grounding tasks.

References

GitHub Hugging Face
This page was created using generative AI such as GPT-5, Claude Opus 4, Gemini 3, Gemini 3.1 Flash Image, and their higher-end successor versions. No guarantee can be made regarding its contents.