Vero: An Open RL Recipe for General Visual Reasoning

A team of researchers has released Vero, a family of fully open vision-language models that match or outperform existing open-weight systems on a broad suite of visual reasoning benchmarks, without relying on proprietary data or reinforcement learning pipelines. The models, detailed in a paper posted to arXiv, are trained using a single-stage reinforcement learning recipe and a curated dataset called Vero-600K, which draws 600,000 samples from 59 distinct datasets [1][2]. The dataset spans six task categories: Chart and OCR, STEM, Spatial and Action, Knowledge and Recognition, Grounding, Counting and Search, and Captioning and Instruction Following [3][4]. The researchers paired this data with task-routed reward functions designed to handle heterogeneous answer formats [2]. When applied to five different starting models, the Vero training protocol yielded average gains of 2.9 to 5.4 points across VeroEval, a 30-benchmark evaluation suite assembled by the authors [1][4]. One variant, Vero-Qwen3I-8B, built on the Qwen3-VL-8B-Instruct model, surpassed the Qwen3-VL-8B-Thinking model by an average of 3.8 points without using additional distillation or proprietary thinking data [1][2]. On the MiMo-VL-SFT base, Vero exceeded MiMo-VL-RL, which itself relies on a proprietary RL recipe [5]. The paper's systematic ablations showed that different task categories produce qualitatively distinct reasoning patterns, and that training on them jointly—rather than in isolation—is necessary for broad performance gains [2][6]. The authors also found that uniform mixture weighting across task categories outperformed weighting schemes based on accuracy, reasoning length, or dataset size [3]. The release includes all training data, code, and model weights. The work was presented as an extended abstract at the 2nd ViSCALE workshop at CVPR 2026 [5]. The authors, including Gabriel Sarch and Zhuang Liu, argue that the results demonstrate broad data coverage is the primary driver of strong RL scaling for visual reasoning [2][6].