Abstract

Reinforcement learning has proven its effectiveness in enhancing thereasoning capabilities of large language models. Recent research efforts haveprogressively extended this paradigm to multimodal reasoning tasks. Due to theinherent complexity and diversity of multimodal tasks, especially in semanticcontent and problem formulations, existing models often exhibit unstableperformance across various domains and difficulty levels. To address theselimitations, we propose VL-Cogito, an advanced multimodal reasoning modeltrained via a novel multi-stage Progressive Curriculum Reinforcement Learning(PCuRL) framework. PCuRL systematically guides the model through tasks ofgradually increasing difficulty, substantially improving its reasoningabilities across diverse multimodal contexts. The framework introduces two keyinnovations: (1) an online difficulty soft weighting mechanism, dynamicallyadjusting training difficulty across successive RL training stages; and (2) adynamic length reward mechanism, which encourages the model to adaptivelyregulate its reasoning path length according to task complexity, thus balancingreasoning efficiency with correctness. Experimental evaluations demonstratethat VL-Cogito consistently matches or surpasses existing reasoning-orientedmodels across mainstream multimodal benchmarks spanning mathematics, science,logic, and general understanding, validating the effectiveness of our approach.