OpenVLThinkerV2: A Generalist Multimodal Reasoning Model for Multi-domain Visual Tasks

2026-04-09

AI Summary

This work addresses the challenges faced by open-source multimodal generalist models in reinforcement learning, particularly the large discrepancies in reward distributions across tasks and the difficulty in jointly optimizing perception and reasoning capabilities. To overcome these issues, the authors propose Gaussian GRPO (G²RPO), a method that enforces alignment of the advantage function to a standard normal distribution via nonlinear distribution matching, thereby ensuring gradient fairness across tasks and enhancing training stability. Additionally, G²RPO incorporates dual shaping mechanisms—response-length shaping and entropy shaping—to dynamically balance visual perception depth with multi-step reasoning exploration. Evaluated on 18 diverse vision-based benchmarks, the proposed approach significantly outperforms both existing open-source and state-of-the-art closed-source models, demonstrating exceptional generalization and robustness.

Abstract

Group Relative Policy Optimization (GRPO) has emerged as the de facto Reinforcement Learning (RL) objective driving recent advancements in Multimodal Large Language Models. However, extending this success to open-source multimodal generalist models remains heavily constrained by two primary challenges: the extreme variance in reward topologies across diverse visual tasks, and the inherent difficulty of balancing fine-grained perception with multi-step reasoning capabilities. To address these issues, we introduce Gaussian GRPO (G$^2$RPO), a novel RL training objective that replaces standard linear scaling with non-linear distributional matching. By mathematically forcing the advantage distribution of any given task to strictly converge to a standard normal distribution, $\mathcal{N}(0,1)$, G$^2$RPO theoretically ensures inter-task gradient equity, mitigates vulnerabilities to heavy-tail outliers, and offers symmetric update for positive and negative rewards. Leveraging the enhanced training stability provided by G$^2$RPO, we introduce two task-level shaping mechanisms to seamlessly balance perception and reasoning. First, response length shaping dynamically elicits extended reasoning chains for complex queries while enforce direct outputs to bolster visual grounding. Second, entropy shaping tightly bounds the model's exploration zone, effectively preventing both entropy collapse and entropy explosion. Integrating these methodologies, we present OpenVLThinkerV2, a highly robust, general-purpose multimodal model. Extensive evaluations across 18 diverse benchmarks demonstrate its superior performance over strong open-source and leading proprietary frontier models.

Problem

• Multimodal Reasoning
• Reinforcement Learning
• Reward Variance
• Perception-Reasoning Balance
• Generalist Models

Innovation

• Gaussian GRPO
• distributional matching
• multimodal reasoning
• entropy shaping
• response length shaping