💬 NLP

Vero: An Open RL Recipe for General Visual Reasoning

Vero is a fully open family of vision-language models that achieves state-of-the-art general visual reasoning performance by scaling reinforcement learning across 600K diverse samples and task-routed rewards, demonstrating that broad data coverage is the primary driver of success without relying on proprietary data or thinking patterns.

Original authors: Gabriel Sarch, Linrong Cai, Qunzhong Wang, Haoyang Wu, Danqi Chen, Zhuang Liu

Published 2026-04-07

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Gabriel Sarch, Linrong Cai, Qunzhong Wang, Haoyang Wu, Danqi Chen, Zhuang Liu

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are trying to teach a robot how to be a genius at looking at pictures and answering questions about them.

Most of the smartest robots out right now (like the ones from big tech companies) are like black boxes. We know they are incredibly smart at solving math problems in pictures or reading charts, but we don't know how they learned it. Their training data and methods are secret recipes locked away in a vault.

This paper introduces Vero, which is like opening the kitchen and showing you the exact recipe to make a world-class visual reasoner. The authors didn't just build one smart robot; they built a whole family of them and released the blueprints, the ingredients, and the cooking instructions to the public.

Here is the simple breakdown of how they did it, using some everyday analogies:

1. The Problem: The "Specialist" Trap

Imagine you hire a chef who is amazing at baking cakes (STEM/math). You then ask them to fix a leaky faucet (spatial reasoning) or write a funny poem about a sunset (creative captioning). They might fail miserably.

The paper found that if you train an AI on just one type of task (like only math), it gets really good at that one thing but actually gets worse at everything else. It's like a musician who only practices scales and forgets how to play a song.

2. The Solution: The "All-You-Can-Eat" Buffet

The authors created a massive dataset called Vero-600K. Think of this as a giant, perfectly balanced buffet with six different food stations:

STEM: Math and science diagrams.
Charts & OCR: Reading graphs and text in images.
Spatial & Action: Figuring out where things are and how to move them (like a robot arm).
Knowledge: Answering questions using what you know about the world.
Grounding & Counting: Finding specific objects and counting them.
Captioning: Writing creative descriptions.

Instead of feeding the robot just cake, they fed it a little bit of everything at the same time. They found that variety is the secret sauce. By eating from all six stations, the robot learned to switch between different "modes" of thinking.

3. The Training Method: The "Coach with a Whistle"

They used a technique called Reinforcement Learning (RL). Imagine a coach training an athlete:

The athlete tries a move.
If they get it right, the coach gives a high-five (a reward).
If they get it wrong, the coach says "try again."

The authors built a smart referee system (called "Task-Routed Rewards").

If the robot is solving a math problem, the referee checks the numbers.
If the robot is writing a poem, the referee checks if it sounds good and follows the rules.
If the robot is pointing at an object, the referee checks if the coordinates are right.

Crucially, they taught the robot to think before it speaks. They forced it to write down its "thought process" (like a scratchpad) before giving the final answer. This is like telling a student, "Show your work," which helps them avoid silly mistakes.

4. The Results: Beating the Secret Giants

The result? Their open-source robot (Vero) became so smart that it beat the "secret" robots from big companies on 23 out of 30 difficult tests.

It could read complex charts better than models trained only on charts.
It could solve spatial puzzles better than models trained only on puzzles.
It could still write creative captions, even though it was also doing hard math.

5. Why This Matters

Before this, if you wanted to build a smart visual AI, you had to guess what data to use or pay a huge company for access.

The "Black Box" Era: "Trust us, our robot is smart because we trained it on secret data."
The Vero Era: "Here is the exact list of 59 datasets we used, here is the code, and here is how we mixed them. You can do it too."

In a nutshell:
The authors proved that to build a truly smart visual AI, you don't need a secret super-recipe. You just need diversity. By mixing different types of visual problems together and giving the AI a clear way to check its own work, you can create a general-purpose "brain" that is just as good as the most expensive, closed-off systems in the world. They handed the keys to the kingdom to everyone.

Technical Summary: Vero: An Open RL Recipe for General Visual Reasoning

1. Problem Statement

Current Vision-Language Models (VLMs) demonstrate strong capabilities in specific visual reasoning domains (e.g., math or OCR), but their performance across a broad spectrum of tasks remains opaque. The most capable models (e.g., GPT-5, Qwen3-VL) rely on proprietary Reinforcement Learning (RL) pipelines with non-public data and undisclosed reward designs. This "black box" nature prevents the research community from:

Systematically diagnosing failure modes.
Understanding which data or reward components drive performance.
Reproducing state-of-the-art (SOTA) results.
Ensuring training safety and transparency.

Furthermore, existing open efforts often focus narrowly on specific domains (e.g., visual math), failing to generalize to other critical capabilities like spatial understanding, grounding, or open-ended instruction following. The core challenge is determining how to train a broadly capable visual reasoner using fully open, transparent methods that avoid the interference and optimization imbalances common in multi-task RL.

2. Methodology

The authors introduce Vero, a family of fully open VLMs trained via a single-stage RL pipeline. The methodology centers on three pillars: Data Curation (Vero-600K), Task-Routed Rewards, and Uniform Mixing.

A. Vero-600K Dataset Construction

The team constructed a 600,000-sample training dataset derived from 59 distinct datasets, organized into six broad task categories to ensure diverse reasoning capabilities:

STEM: Mathematical diagrams, scientific figures, medical imaging.
Spatial & Action: Embodied reasoning, UI navigation, 3D spatial understanding.
Knowledge & Recognition: VQA combining object recognition with commonsense knowledge.
Chart & OCR: Extracting and reasoning over structured data (charts, tables, infographics).
Grounding, Counting & Search: Bounding box localization, counting, and visual search.
Captioning & Instruction Following: Open-ended description and prompt adherence.

Curation Pipeline:

Sourcing: Started with >250 candidate datasets.
Heuristic Filtering: Removed datasets with <1k samples, low resolution, or binary questions.
Manual QC: Inspected ~50 samples per dataset for correctness, unambiguity, and verifiability.
LLM-Based Filtering: Used a strong VLM (Qwen3-VL-235B) to filter ambiguous questions, image-irrelevant prompts, and unverifiable answers.
Answer Normalization: Standardized ground truths (e.g., stripping units, converting fractions to decimals) to ensure stable reward computation.

B. Task-Routed Reward Design

A critical innovation is the multi-task reward system. Instead of a single reward function, the system routes the answer through specific verifiers based on the task type:

String Match: Exact equality for text.
Multiple Choice: Letter extraction and comparison.
Numeric: Symbolic parsing via MATH-VERIFY with tolerance.
Grounding/Clicking: IoU/F1 scoring for bounding boxes or point coordinates.
Web Action: Weighted field matching for JSON actions.
Instruction Following: Programmatic constraint checking.
Open-Ended: LLM-as-a-Judge (using Qwen3-32B) to score responses against reference answers, with strict guidelines to prevent "reward hacking" (e.g., penalizing self-evaluative language).

The total reward combines accuracy ( $R_{acc}$ ), format compliance ( $R_{fmt}$ ), and an overlong penalty ( $R_{overlong}$ ).

C. Training Algorithm

Algorithm: The models are trained using GSPO (Group Sequence Policy Optimization), an improvement over GRPO that uses sequence-level importance ratios to stabilize training.
Mixing Strategy: The authors found that uniform mixing of the six task categories (100k samples each) outperforms strategies weighted by difficulty, dataset size, or reasoning length.
Base Models: The recipe was applied to four base models: Qwen2.5-VL-7B, Qwen3-VL-8B (Instruct & Thinking), and MiMo-VL-7B.

3. Key Contributions

Fully Open SOTA Recipe: Vero provides the first fully open, reproducible RL recipe that matches or exceeds proprietary models (like Qwen3-VL-Thinking) across diverse visual tasks without using proprietary data.
Vero-600K & VeroEval: The release of a 600k-sample multi-task dataset and a comprehensive 30-benchmark evaluation suite (VeroEval) covering six distinct reasoning categories.
Discovery of Task-Specific Reasoning Patterns: Through systematic ablations, the authors demonstrate that different task categories elicit qualitatively distinct reasoning behaviors (e.g., STEM triggers backtracking; Grounding suppresses introspection in favor of directed search).
Negative Transfer in Single-Task Training: The study proves that training on a single task category often leads to negative transfer on other tasks (e.g., training on STEM degrades Captioning performance). Broad data coverage is essential for generalization.
Reward Design Insights: Demonstrated that simple, unified rewards (like math_verify) fail on heterogeneous tasks, while task-routed rewards are necessary for multi-task success.

4. Experimental Results

Performance on VeroEval (30 Benchmarks)

Overall SOTA: Vero-Qwen3T-8B (trained on Qwen3-VL-8B-Thinking) achieved an average score of 65.9, outperforming the base Qwen3-VL-8B-Thinking on 24 of 30 benchmarks.
Beating Proprietary Models: Vero-Qwen3I-8B (trained on Qwen3-VL-8B-Instruct) outperformed Qwen3-VL-8B-Thinking on 23 of 30 benchmarks, despite the latter having access to proprietary "thinking" data.
Cross-Model Gains: Training on Vero-600K improved four different base models by 3.7 to 5.5 points on average. Notably, Vero-Qwen25-7B (starting from a weaker 7B model) surpassed Qwen3-VL-8B-Instruct on Chart & OCR and Knowledge & Recognition categories.
Comparison to Open RL: Vero-600K significantly outperformed existing open RL datasets (ViRL-39k, OpenMMReasoner-74k, LLaVA-OV-1.5-RL-70k) across almost all categories, particularly in Grounding, Counting & Search (+5.9 points) and Captioning (+15.0 points).

Ablation Findings

Data Diversity: Uniform mixing of all six categories yielded the best overall performance (+5.8 points). Weighting by difficulty or reasoning length improved specific categories but hurt others.
Single-Task Failure: Models trained on single categories suffered catastrophic drops in non-target tasks (e.g., Captioning-only training reduced performance by -15 to -35 points on other tasks).
Reasoning Length: Different tasks induce different reasoning depths. Spatial & Action tasks required the longest reasoning (avg. ~~1983 words), while Knowledge & Recognition was shortest (~~76 words), suggesting distinct cognitive regimes.

5. Significance and Impact

Democratization of RL: Vero proves that SOTA visual reasoning does not require proprietary "secret sauce" data. It shifts the paradigm from "black box" scaling to transparent, data-centric engineering.
Understanding Generalization: The paper provides empirical evidence that breadth of task coverage is the primary driver of general visual reasoning, more so than algorithmic sophistication. It challenges the notion that models can simply "learn to think" without diverse behavioral exposure.
Safety and Reproducibility: By releasing all data, code, and models, the authors enable the community to audit training procedures, verify safety, and build upon a solid foundation for future multimodal research.
Cognitive Insights: The findings align with cognitive science, suggesting that VLMs, like humans, utilize distinct "task sets" and reasoning strategies for different problems, and that a generalist model must be trained to switch between these modes effectively.

In conclusion, Vero establishes a new benchmark for open-source visual reasoning, demonstrating that a carefully curated, diverse dataset combined with task-aware reward design is sufficient to train models that rival or exceed proprietary systems across a wide array of complex visual tasks.