💻 computer science

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

The paper proposes Perceive-to-Reason (P2R), a unified framework that decouples fine-grained visual reasoning into distinct perception and reasoning stages, enhanced by a role-aware reinforcement learning strategy (PRA-GRPO) to significantly boost performance across various model scales and high-resolution benchmarks.

Original 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

Published 2026-07-02

📖 4 min read☕ Coffee break read

CC BY 4.0

Original 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

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are trying to solve a very tricky puzzle, but the picture is huge and the most important clue is a tiny speck hidden somewhere in the middle of a crowded scene.

This is the problem Perceive-to-Reason (P2R) tries to solve for Artificial Intelligence.

The Problem: The "Needle in a Haystack"

Current AI models (Vision-Language Models) are great at looking at a picture and giving a general answer. But when the question requires finding a tiny detail—like reading a small sign in a high-resolution photo or spotting a specific color on a distant object—they often get lost.

Think of it like trying to find a specific friend in a stadium full of 50,000 people.

Old AI methods are like someone who just stares at the whole crowd and guesses, "I think they are over there!" or frantically zooms in and out of random sections without a plan. They often miss the tiny details because they try to do everything at once.
The Paper's Insight: The researchers realized the AI was failing not because it couldn't think (reason), but because it couldn't look (perceive) correctly first. It was trying to solve the math problem before it even found the numbers.

The Solution: The Detective and The Judge

The paper proposes a new way to train AI called Perceive-to-Reason. Instead of making the AI do everything in one big brain burst, they split the job into two distinct roles, like a team of two people:

The Perceiver (The Detective):
- Job: Look at the huge, high-resolution image and the question. Ignore the answer for a moment. Just find the specific spot where the clue is hiding.
- Action: It draws a box around the relevant area (e.g., "The red chair is here") and crops that small piece out.
- Analogy: This is like a detective scanning a crime scene, finding the single muddy footprint, and putting it in an evidence bag.
The Reasoner (The Judge):
- Job: Look only at the evidence bag (the cropped image) and the original picture with the box drawn on it. Now, answer the question.
- Action: It uses the clear, zoomed-in view to make the final decision.
- Analogy: This is the judge looking only at the muddy footprint to decide if it matches the suspect's shoe. They don't have to worry about the rest of the messy crime scene anymore.

The Secret Sauce: "PRA-GRPO"

You might ask, "How do you teach an AI to do this if you only tell it 'Right' or 'Wrong' at the very end?" (You don't have a teacher pointing out exactly where the box should go).

The authors invented a training method called PRA-GRPO. Think of this as a specialized coaching drill for the AI team:

The Drill: The AI practices in alternating rounds.
- Round 1 (Perception Focus): The "Detective" tries to find the spot. The "Judge" is frozen (not learning). If the Detective finds the right spot, the Judge can easily get the answer right. The AI gets a "good job" signal for the Detective.
- Round 2 (Reasoning Focus): The "Detective" is frozen (using what it just learned). The "Judge" tries to answer based on that spot. If the Judge gets it right, it gets a "good job" signal.
The Result: By taking turns, the AI learns that a good Detective makes the Judge's job easier, and a good Judge gives better feedback to the Detective. They learn to work together without needing a human to draw the boxes for them.

What Happened?

The researchers tested this on models of different sizes (2 billion, 4 billion, and 8 billion "brain cells").

The Result: The new method (P2R) was significantly better at finding tiny details and answering complex questions than the original models.
The Proof: On a test called V-Star, the 4-billion model jumped from getting 81.7% of answers right to 93.2%. That's a massive improvement, especially for tasks involving high-resolution images where details are small.

Summary

In simple terms, this paper says: "Don't try to think and look at the same time. First, find the right piece of the picture. Then, think about it."

By separating the act of finding the clue from the act of solving the puzzle, and training the AI to take turns practicing these skills, the computer becomes much better at seeing the small things that matter.

Technical Summary: Perceive-to-Reason (P2R)

Problem Statement

Fine-grained visual reasoning remains a significant bottleneck for Vision-Language Models (VLMs), particularly when critical visual cues are subtle and buried within high-resolution images. Existing approaches typically rely on "Thinking with Images" (interleaving region exploration with reasoning) or "Visual Search" (test-time search for key regions). However, these methods often entangle perception (locating evidence) with reasoning (answering the question) or depend on complex, difficult-to-optimize external pipelines. A diagnostic study presented in the paper indicates that a primary cause of failure is not a lack of reasoning capability, but the inability to correctly localize the relevant visual evidence in the first place. Furthermore, training decoupled perception and reasoning stages is challenging under standard answer-only supervision, as it is difficult to attribute errors to either the localization step or the reasoning step.

Methodology

The authors propose Perceive-to-Reason (P2R), a unified framework that explicitly decouples fine-grained visual reasoning into two distinct stages: Perception and Reasoning.

1. Two-Stage Inference Pipeline

P2R utilizes a single underlying VLM with shared parameters but operates in two consecutive roles:

Perceiver Stage: Given an image $I$ and a question $Q$ , the model acts as a Perceiver to explicitly localize question-relevant evidence. It outputs bounding boxes $B$ representing key regions. These boxes are used to generate an annotated image (with boxes overlaid) and cropped evidence images.
Reasoner Stage: The model then acts as a Reasoner, taking the original question, the annotated image, and the cropped regions as input to generate the final answer.

This formulation transforms fine-grained reasoning into a structured process where the model first determines where to look and then determines how to reason based on that evidence.

2. Perception-Reasoning Alternating GRPO (PRA-GRPO)

To train this decoupled framework using only final-answer supervision (without ground-truth bounding box annotations), the authors introduce PRA-GRPO, a role-aware reinforcement learning strategy based on Group Relative Policy Optimization (GRPO).

Alternating Optimization: The training process alternates between optimizing the Perceiver and the Reasoner. In the Perception Phase, the Perceiver is updated while the Reasoner is frozen; the reward is derived from the final answer correctness generated by the frozen Reasoner. Conversely, in the Reasoning Phase, the Reasoner is updated while the Perceiver is frozen.
Attributable Signals: This approach converts the binary final-answer reward into a role-specific learning signal. It allows the model to learn effective evidence localization based on downstream reasoning outcomes, effectively solving the credit assignment problem in a decoupled pipeline without requiring dense supervision.
Shared Parameters: Both roles share the same model parameters $\theta$ , allowing the capabilities to be integrated within a single model instance.

Key Contributions

P2R Framework: A unified framework that formulates fine-grained visual reasoning as a two-stage perceive-to-reason process, explicitly separating evidence localization from answer generation.
PRA-GRPO: A novel training strategy that aligns reinforcement learning with the decoupled formulation. It enables the learning of perception capabilities using only final-answer supervision, eliminating the need for ground-truth bounding box annotations.
State-of-the-Art Performance: The method demonstrates consistent improvements across model scales (2B, 4B, 8B) and benchmarks, achieving new state-of-the-art results on high-resolution fine-grained visual reasoning tasks.

Experimental Results

The authors evaluated P2R on several benchmarks, including V-Star, HR-Bench-4K/8K, and MME-RealWorld-Lite.

High-Resolution Benchmarks: P2R consistently outperforms its base models (Qwen3-VL-Instruct) and existing visual search or "Thinking with Images" methods.
- P2R-4B achieved 93.2% on V-Star, 81.9% on HR-Bench-4K, and 80.5% on HR-Bench-8K.
- Compared to the Qwen3-VL-Instruct-4B backbone, P2R-4B showed gains of +11.5% on V-Star, +8.1% on HR-Bench-4K, and +13.5% on HR-Bench-8K.
Generalization: The benefits of P2R extend beyond high-resolution tasks. On the MME-RealWorld-Lite benchmark, P2R improved performance across diverse real-world multimodal reasoning tasks (OCR, remote sensing, diagrams), suggesting the perceive-to-reason formulation aids broader multimodal understanding.
Ablation Studies:
- Inference: Replacing direct Chain-of-Thought (CoT) with the P2R inference pipeline significantly improved performance even on base models.
- Training: Alternating updates between roles (P $\to$ R) yielded better results than training roles in isolation or in reverse order, validating the importance of the decoupled training schedule.
- Parameter Sharing: Experiments showed that training a role (e.g., Perceiver) and reusing those weights for the other role (Reasoner) provided benefits, though the fully alternating PRA-GRPO training yielded the best results.

Significance and Claims

The paper claims that explicitly decoupling perception from reasoning provides an effective framework for fine-grained visual reasoning. By treating evidence localization as an explicit intermediate step rather than an implicit byproduct of answer generation, P2R addresses the primary bottleneck of finding the right visual evidence.

The significance of PRA-GRPO lies in its ability to train this decoupled process using only final-answer supervision. This removes the dependency on expensive, task-specific bounding box annotations while still enabling the model to learn precise localization. The authors conclude that their approach offers a scalable and effective solution for high-resolution visual reasoning, achieving substantial gains over existing baselines and demonstrating that the benefits of this formulation extend to broader multimodal reasoning scenarios.

Limitations Acknowledged by the Authors:

The two-stage pipeline introduces additional inference latency compared to direct prompting.
The full scaling behavior of PRA-GRPO at larger training scales was not explored due to computational constraints.
The reliance on final-answer supervision provides a sparse learning signal.
Evaluation is currently focused on fine-grained reasoning and related benchmarks, with generalization to interactive or long-horizon settings remaining for future work.