Enterprise teams building AI-powered document analysis, automated inspection, or any visual understanding module have long faced a dilemma: the most capable vision-language models (VLMs) come from closed pipelines whose training data, reward formulas, and reinforcement learning (RL) recipes remain proprietary. This opacity makes it difficult to audit, customize, or reproduce their performance in production environments. A new research paper published on arXiv titled Vero: An Open RL Recipe for General Visual Reasoning directly addresses that problem by releasing a fully open family of VLMs that match or exceed existing open-weight models across diverse visual reasoning tasks.
Open Recipe for Visual Reasoning
The research team—comprising Sarch, Gabriel, Cai, Linrong, Wang, Qunzhong, Wu, Haoyang, Chen, Danqi, Liu, and Zhuang—designed Vero as a “fully open” alternative to state-of-the-art but closed VLMs. According to the paper, the strongest vision-language models “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.” Vero aims to lift that barrier by publishing all data, code, and model weights.
Dataset and Reward Design
To train Vero, the team constructed Vero-600K, a dataset of 600,000 samples drawn from 59 different datasets covering six broad task categories: charts, science, spatial understanding, and open-ended tasks among others. A key innovation is the use of “task-routed rewards” that handle the heterogeneous answer formats across these diverse tasks—multiple choice, free text, numerical values—without forcing a one-size-fits-all scoring function. The paper describes this as a way to scale RL data and rewards across the six categories.
Performance Benchmarks
The team evaluated Vero on VeroEval, a suite of 30 benchmarks. Controlled comparisons showed that Vero-600K outperforms existing RL datasets. When applied to five different starting models, Vero variants achieved average gains of 2.9 to 5.4 points over their initial models. Notably, the variant Vero-Qwen3I-8B—trained on the Instruct model—surpassed Qwen3-VL-8B-Thinking by 3.8 points on average without any additional distillation. The following table summarizes the key performance improvements:
| Model Variant | Average Gain Over Baseline |
|---|---|
| All Vero variants (range across 5 starting models) | 2.9–5.4 points |
| Vero-Qwen3I-8B vs Qwen3-VL-8B-Thinking | 3.8 points |
Ablations and Joint Learning
Systematic ablations revealed that different task categories elicit distinct reasoning patterns and that broad gains depend on learning these patterns jointly rather than in isolation. According to the paper, “broad gains depend on learning them jointly rather than in isolation.” This finding has practical implications for enterprise teams that may want to fine-tune models for a specific domain: a narrow focus on a single task type may limit overall reasoning capability.
Implications for Custom AI
Because all data, code, and models are publicly available, technology leaders in regulated industries or niche verticals can audit the training pipeline, adapt the reward structure to their own data, and fine-tune Vero without relying on a vendor’s black box. While the paper does not target supply chain or logistics directly, the ability to build custom visual reasoners on open technology directly supports use cases such as automated document reading, quality inspection, and geospatial analysis. The Vero recipe provides a reproducible foundation that enterprises can extend, rather than a finished product locked behind a proprietary API.