Vero: An Open RL Recipe for General Visual Reasoning
Vero is an open RL method for visual reasoning, using the Vero-600K dataset to enhance performance.
Key Findings
Methodology
Vero employs the Vero-600K dataset, integrating six task categories with task-routed reward mechanisms to enhance visual reasoning. It uses the GRPO algorithm to optimize multi-task learning, ensuring superior performance across diverse visual tasks.
Key Results
- Vero-600K outperforms existing RL datasets across 30 benchmarks, with an average improvement of 2.9 to 5.4 points.
- Vero-Qwen3I-8B surpasses Qwen3-VL-8B-Thinking by 3.8 points without additional distillation.
- Different task categories elicit distinct reasoning patterns, with joint learning outperforming isolated learning.
Significance
Vero provides a fully open platform for visual reasoning research, addressing the reproducibility challenges of existing closed data and RL pipelines. Its success in multi-task learning offers new research directions for academia and industry.
Technical Contribution
Vero significantly enhances visual reasoning model performance through task-routed rewards and multi-task dataset design, providing new theoretical guarantees and engineering possibilities.
Novelty
Vero is the first RL method to achieve multi-task visual reasoning in an open environment, overcoming the limitations of existing closed-source methods.
Limitations
- Vero still faces performance instability in certain task categories, especially under imbalanced datasets.
- The model may underperform in handling extremely complex visual scenes.
Future Work
Future work includes expanding the Vero-600K dataset, exploring more task categories, and optimizing task-routed reward mechanisms to improve model generalization.
AI Executive Summary
Vero is an open-source reinforcement learning method for visual reasoning, addressing the reproducibility challenges of existing vision-language models in multi-task reasoning. By constructing the Vero-600K dataset, Vero excels across six task categories, significantly enhancing reasoning capabilities.
The methodology of Vero is based on the GRPO algorithm, combined with task-routed reward mechanisms, ensuring superior performance across diverse visual tasks. Experimental results show that Vero outperforms existing RL datasets across 30 benchmarks, with an average improvement of 2.9 to 5.4 points.
Vero provides a fully open platform for visual reasoning research, addressing the reproducibility challenges of existing closed data and RL pipelines. Its success in multi-task learning offers new research directions for academia and industry.
Deep Dive
Abstract
What does it take to build a visual reasoner that works across charts, science, spatial understanding, and open-ended tasks? The strongest vision-language models (VLMs) suggest that broad visual reasoning is within reach, yet their closed data and reinforcement learning (RL) pipelines make their gains difficult to study, reproduce, or extend. We introduce Vero, a family of fully open VLMs that match or exceed existing open-weight models across diverse visual reasoning tasks. We scale RL data and rewards across six broad task categories, constructing Vero-600K, a 600K-sample dataset from 59 datasets, and designing task-routed rewards that handle heterogeneous answers. Across VeroEval, our 30-benchmark suite, Vero-600K outperforms existing RL datasets under controlled comparisons. Applied to five starting models, Vero variants gain 2.9-5.4 points on average over their initial models. Notably, Vero-Qwen3I-8B, trained on the Instruct model, surpasses Qwen3-VL-8B-Thinking by 3.8 points on average without additional distillation. Systematic ablations reveal that different task categories elicit distinct reasoning patterns and that broad gains depend on learning them jointly rather than in isolation. All data, code, and models are publicly available.