Vero: An Open RL Recipe for General Visual Reasoning
| Authors | Gabriel Sarch et al. |
| Year | 2026 |
| HF Upvotes | 30 |
| arXiv | 2604.04917 |
| Download | |
| HF Page | View on Hugging Face |
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) show such broad visual reasoning is within reach, but the recipe behind them remains unclear, locked behind proprietary reinforcement learning (RL) pipelines with non-public data. We introduce Vero, a family of fully open VLMs that matches or exceeds 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 answer formats. Vero achieves state-of-the-art performance, improving over four base models by 3.7-5.5 points on average across VeroEval, our suite of 30 challenging benchmarks. Starting from Qwen3-VL-8B-Instruct, Vero outperforms Qwen3-VL-8B-Thinking on 23 of 30 benchmarks without additional proprietary thinking data. When trained from the same base model, Vero-600K exceeds existing RL datasets across task categories. Systematic ablations reveal that different task categories elicit qualitatively distinct reasoning patterns that transfer poorly in isolation, suggesting that broad data coverage is the primary driver of strong RL scaling. All data, code, and models are released.
Engineering Breakdown
Plain English
Vero is an open-source family of vision-language models (VLMs) that achieves state-of-the-art performance on diverse visual reasoning tasks—spanning charts, scientific imagery, spatial understanding, and open-ended problems—by combining reinforcement learning with a large-scale, publicly available dataset. The authors created Vero-600K, a 600,000-sample dataset aggregated from 59 existing datasets across six task categories, and designed task-routed rewards that handle heterogeneous answer formats (multiple choice, free-form text, numerical, etc.). Across four different base models, Vero improved performance by 3.7–5.5 points on average, matching or exceeding existing closed-source proprietary models while remaining fully transparent and reproducible.
Core Technical Contribution
The core contribution is demonstrating that broad visual reasoning capability can be unlocked through a combination of: (1) scaling RL training data to 600K samples across six heterogeneous task categories, (2) designing task-routed reward models that adapt to different answer formats rather than forcing a single reward signal, and (3) making the entire pipeline—data, rewards, and model weights—publicly available. Prior work either kept recipes proprietary or worked with smaller, more domain-specific datasets; Vero shows that open, large-scale multi-task RL on diverse visual reasoning benchmarks produces competitive results without trade secrets. This directly challenges the assumption that strong visual reasoning requires closed proprietary training infrastructure.
How It Works
Vero operates as a multi-stage reinforcement learning pipeline applied to vision-language models. First, a base VLM (e.g., Claude, LLaVA, or others) is used as the starting point. Second, the system ingests Vero-600K, a composite dataset built by merging 59 different visual reasoning benchmarks—covering charts, diagrams, scientific images, spatial reasoning, and general knowledge questions. Third, task-routed rewards are applied: instead of one generic reward model, the system routes each task to a specialized reward function that understands its particular output format (e.g., a reward for multiple-choice differs from one for free-form text). Fourth, the base model is fine-tuned via RL using these task-specific rewards, optimizing the policy to maximize task-appropriate signals. The resulting model outputs improved visual reasoning across all six categories without losing generality.
Production Impact
For engineers building visual AI systems, Vero offers a reproducible open-source recipe to improve VLM performance by 3–5 points on real benchmarks without proprietary data or compute. Rather than licensing closed models or reverse-engineering reward signals, teams can directly use Vero's 600K dataset and task-routed reward designs to fine-tune models in-house, reducing vendor lock-in and enabling customization for domain-specific charts or images. The multi-task, task-routed reward approach is directly applicable to production fine-tuning pipelines: if your deployment spans multiple reasoning domains (e.g., OCR, tables, diagrams, and spatial reasoning), routing rewards by task type reduces reward model brittleness and improves convergence. Trade-offs include computational cost (RL fine-tuning is expensive; expect 2–10x training time vs. supervised fine-tuning), the need to curate or validate task-specific reward models for new domains, and potential distribution shift if production images differ substantially from the 59 aggregated datasets.
Limitations and When Not to Use This
Vero's 600K-sample aggregate dataset, while large, is drawn entirely from existing public benchmarks and may encode their annotation biases and domain coverage gaps—it does not show how to handle truly novel visual reasoning tasks outside the six categories. The paper does not discuss performance degradation on out-of-distribution images (e.g., real-world photos, proprietary diagrams, or adversarial visual inputs), nor does it provide guidance on when to add task-routed rewards vs. a unified reward signal for new domains. Scaling and data efficiency are not deeply analyzed; it remains unclear whether the 3.7–5.5 point gains scale logarithmically with dataset size or if diminishing returns appear earlier, and no ablation compares the cost of task-routed vs. simpler unified reward models. Finally, the paper assumes that existing base VLMs are strong enough to benefit from RL on these tasks; it does not address how to apply Vero to smaller or more efficient models, which may have different reward learning dynamics.
Research Context
Vero builds on a decade of work in vision-language modeling (CLIP, ALIGN, LLaVA, GPT-4V) and recent advances in RL-based fine-tuning (RLHF for language models, RLAIF). It directly responds to growing interest in making multimodal reasoning reproducible and open—contrasting with proprietary systems like GPT-4V and Claude Vision. The paper aggregates and systematizes visual reasoning benchmarks (MMVet, ChartQA, ScienceQA, etc.) into a unified RL framework, advancing the field's understanding of how to combine heterogeneous tasks under one training signal. This work opens a research direction toward truly general visual reasoning by demonstrating that scale and task routing, not proprietary secrets, may be the primary drivers of broad capability.
:::tip Subscribe Get weekly breakdowns of papers like this in AI Letters - the newsletter for engineers building production AI systems. :::