Vero: An Open RL Recipe for General Visual Reasoning

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 to 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.

State of the Art Performance Across Task Categories

We evaluate Vero on 30 benchmarks spanning six task categories. The same open recipe improves six different initial models and achieves strong overall performance across model families.

Vero-Qwen35-9B reaches 74.4 overall and improves over Qwen3.5-9B by +2.9, with gains on 25 of 30 benchmarks and all six category averages. Applied directly to Qwen3.5-9B-Base, Vero-Qwen35-9B-Base reaches 73.0 overall, a +12.9-point gain without an SFT or distillation warm start. Among 8B models, Vero-Q3I-8B reaches 66.1 overall and outperforms Qwen3-VL-8B-Thinking by +3.8 overall, while Vero-Q3T-8B reaches 65.8 overall and improves over Qwen3-VL-8B-Thinking by +3.5. The same recipe improves Qwen2.5-VL-7B-Instruct from 52.9 to 57.8 and MiMo-VL-7B-SFT to 63.2, exceeding MiMo-VL-7B-RL at 62.4.

† indicates evaluated by us. All other results are taken from official reports.

Vero Demos

Example conversations between a user and Vero across all six task categories. Each demo shows the model's reasoning trace and final answer.

Vero-Q3T-8B

Method

Vero trains on 600K curated RL samples drawn from 59 datasets organized into six categories: Chart and OCR, STEM, Spatial and Action, Knowledge and Recognition, Grounding, Counting and Search, and Captioning and Instruction Following. The categories correspond to substantially different use cases, visual inputs, reasoning patterns, and answer formats.

The training mixture spans six task categories and 59 retained datasets after dataset-level and sample-level filtering.

Vero uses a single-stage RL recipe directly on top of instruction-tuned or RL'd base models. Vero uses GSPO-style optimization with task-routed verifiers, so numeric questions, multiple choice questions, grounding boxes, clicks, ordering problems, and open-ended instruction-following outputs.

Data Diversity and Transfer

We show that single-task RL does not generalize reliably across visual capabilities. Training on one category often improves that category while degrading others, especially Grounding and Captioning and Instruction Following. This is consistent with classic multi-task RL results showing that heterogeneous tasks can interfere and that task contributions must be balanced during training (Teh et al., 2017; Hessel et al., 2019). By contrast, the mixed model produces positive gains across categories and avoids the catastrophic spillover seen in single-task-category RL.

Behavioral Analysis

Different task categories do not simply induce more or less reasoning — they induce qualitatively different reasoning styles. STEM tasks trigger reflective, backtracking-heavy traces; grounding tasks favor direct perceptual search; chart tasks produce systematic regional synthesis. These distinct patterns help explain why single-task training transfers poorly: the model adapts not just its answers, but its reasoning policy.

Reasoning Length by Task Category

Beyond qualitative differences in reasoning style, task categories also elicit markedly different reasoning lengths. Spatial & Action produces the longest responses at 1,983 ± 51 words, followed by Chart & OCR (1,593 ± 32) and STEM (1,576 ± 40). Captioning & Instruction Following is much shorter (414 ± 13), while Grounding, Counting & Search (125 ± 13) and Knowledge & Recognition (76 ± 3) are shortest. The gap between the longest and shortest categories exceeds 26×, suggesting that long chain-of-thought behavior is concentrated in tasks requiring multi-step spatial state tracking or structured analytical decomposition.

RL on different task categories leads to varying reasoning lengths. Average reasoning length (in words) on the validation set, measured after training Qwen3-VL-8B-Instruct for 1,000 steps on each task category (100k samples) and evaluating on the same category. Error bars denote the standard error of the mean.

Task categories cultivate distinct skill repertoires. A logistic regression probe trained on 1,500 extracted skills per task category reaches high overall accuracy at recovering the source task category from the extracted skill lists, suggesting partially distinct skill repertoires rather than a generic increase in chain-of-thought length.

Interactive UMAP

The stacked-bar summary highlights the same task category separability at the category level. The interactive UMAP below shows the same story at the individual-skill level, where clusters can be inspected directly by task category, label, and description.

Citation

If you find Vero useful in your research, please consider citing:

@inproceedings{sarch2026vero,
  title={Vero: An Open RL Recipe for General Visual Reasoning},
  author={Sarch, Gabriel and Cai, Linrong and Wang, Qunzhong and Wu, Haoyang and Chen, Danqi and Liu, Zhuang},
  booktitle={ECCV},
  year={2026}
}