Vero: An Open RL Recipe for General Visual Reasoning

2025-08-07 · Source: cs.CL updates on arXiv.org · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Computer Vision · Depth: Expert, extended

Summary

Vero is a new family of fully open vision-language models (VLMs) designed for general visual reasoning, matching or exceeding existing open-weight models. Developed using an open reinforcement learning (RL) recipe, Vero leverages Vero-600K, a 600,000-sample dataset compiled from 59 datasets across six broad task categories, alongside task-routed reward functions. This approach yields strong performance, improving over four base models by 3.7–5.5 points on average across VeroEval, a suite of 30 challenging benchmarks. For instance, Vero-Qwen3T-8B surpassed Qwen3-VL-8B-Thinking on 24 of 30 benchmarks, with notable gains such as +12.1 on ScreenSpotPro. Ablations confirm that broad data coverage, uniform task category weighting, and task-specific reward design are critical for robust RL scaling and preserving visual chat capabilities. All data, code, and models are publicly released.

Key takeaway

For machine learning engineers developing general visual reasoning VLMs, you should prioritize diverse, multi-task training data and task-specific reward functions. Vero's open recipe demonstrates that broad data coverage, uniform category mixing, and routed rewards are critical for achieving high performance and preventing reasoning blind spots. Consider adopting the Vero-600K dataset and VeroEval suite to enhance your model's capabilities across varied visual tasks.

Key insights

Broad data diversity and task-routed rewards are crucial for general visual reasoning in open VLMs.

Principles

• Broad data coverage drives strong RL scaling for VLMs.
• Uniform task category weighting optimizes multi-task VLM training.
• Task-routed reward design is essential for heterogeneous answer formats.

Method

Vero's single-stage RL recipe uses Vero-600K (59 datasets, 6 categories) with task-routed reward functions. It involves multi-stage data filtering and uniform task category mixing.

In practice

• Use Vero-600K dataset for VLM RL training.
• Implement task-routed rewards for diverse visual tasks.
• Evaluate with VeroEval's 30 benchmarks.

Topics

• Vision-Language Models
• Reinforcement Learning
• Visual Reasoning
• Multi-task Learning
• Vero-600K Dataset
• VeroEval Benchmark
• Reward Design

Code references

• zlab-princeton/vero

Best for: Research Scientist, AI Engineer, Computer Vision Engineer, AI Scientist, Machine Learning Engineer, AI Architect

Related on AIssential

• Reward Design for Physical Reasoning in Vision-Language Models — Reward design significantly shapes VLM physical reasoning, with accuracy and attention-based signals driving domain-specific performance gains.
• The Death of RLHF: A Practitioner’s Guide to the New Post-Training Stack — New post-training methods like GRPO, RLVR, and DAPO have replaced RLHF for reasoning tasks, improving scalability and efficiency.
• STRIDE: Strategic Trajectory Reasoning via Discriminative Estimation for Verifiable Reinforcement Learning — STRIDE enhances RLVR by using verifiable outcomes to identify and reward outcome-discriminative strategic patterns, improving credit assignment.
• Decompose, Look, and Reason: Reinforced Latent Reasoning for VLMs — DLR enhances VLM visual reasoning by dynamically decomposing queries and extracting premise-conditioned continuous visual latents.
• RLVR without Ineffective Samples: Group Prioritized Off-Policy Optimization for LLM Reasoning — POPO improves LLM reasoning by efficiently leveraging effective off-policy data in RLVR, reducing computational overhead.
• CAVE: A Structured Credit Assignment Approach for Fragmented Visual Evidence Reasoning — CAVE improves VLM reasoning by assigning structured, action-level credits for visual evidence exploration.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by cs.CL updates on arXiv.org.