Seeing with You: Perception-Reasoning Coevolution for Multimodal Reasoning

📰 ArXiv cs.AI

Researchers propose a perception-reasoning coevolution approach for multimodal reasoning in large language models

advanced Published 31 Mar 2026

Action Steps

Identify the limitations of existing reinforcement learning with verifiable rewards (RLVR) approaches
Develop a perception-reasoning coevolution framework that updates perception and reasoning separately
Implement the framework using multimodal large language models (MLLMs) and evaluate its performance
Analyze the results and refine the framework to improve multimodal reasoning capabilities

Who Needs to Know This

AI researchers and engineers working on multimodal large language models can benefit from this approach to improve reasoning capabilities, and software engineers can apply this to develop more advanced AI models

Key Insight

💡 Separate updates for perception and reasoning can improve credit assignment and enhance multimodal reasoning capabilities

Key Takeaways

Researchers propose a perception-reasoning coevolution approach for multimodal reasoning in large language models

Full Article

Title: Seeing with You: Perception-Reasoning Coevolution for Multimodal Reasoning

Abstract:
arXiv:2603.28618v1 Announce Type: new Abstract: Reinforcement learning with verifiable rewards (RLVR) has substantially enhanced the reasoning capabilities of multimodal large language models (MLLMs). However, existing RLVR approaches typically rely on outcome-driven optimization that updates both perception and reasoning using a shared reward based solely on the final answer. This shared reward blurs credit assignment, frequently improving reasoning patterns while failing to reliably enhance th

Read full paper → ← Back to Reads