---
title: "VBVR-Pro: A Scalable and Verifiable Suite for Native Visual Reasoning"
source: "arXiv AI/CL/LG（论文）"
source_tier: "t2"
url: "https://arxiv.org/abs/2608.26105v1"
permalink: "https://feed.astralor.com/items/itm_615e5519323f532c262da8b4"
published: "2026-08-26T17:59:51.000Z"
score: 78
channel: "ai"
tags: ["VBVR-Pro","Native Visual Reasoning","Verifiable Reward","Reinforcement Learning","Visual Generation"]
---

# VBVR-Pro: A Scalable and Verifiable Suite for Native Visual Reasoning

## AI commentary

> This paper turns “generation-as-reasoning” from a slogan into an experimental system, with task scaling, reliable rewards, and modality ablations all in one package. It is especially valuable for researchers working on visual RL, multimodal evaluation, and generative reasoning, and the released assets could become a new community baseline.

## At a glance

VBVR-Pro is a closed-loop testbed for native visual reasoning, treating visual generation itself as the medium of reasoning. It features 300 procedurally generated tasks that enable scalable training, with models showing strong transfer to seven external benchmarks including RISE-Video, MME-CoF-Pro, and BabyVision. It provides verifiable reward scorers grounded in deterministic task-specific rules, overcoming common failure modes of the VLM-as-a-judge paradigm and serving as reliable reward signals for large-scale multi-task reinforcement learning. Furthermore, it enables controlled comparisons across over 30 image, video, and interleaved generators, showing that video generation is strongest for persistent spatiotemporal state tracking while interleaved generation is a compute-efficient alternative. All data, models, scorers, and code are released.

## Key points

- A controlled task space of 300 procedurally generated tasks enables scalable training, with strong transfer to seven external benchmarks such as RISE-Video, MME-CoF-Pro, and BabyVision.
- Verifiable reward scorers grounded in deterministic rules outperform the prevalent VLM-as-a-judge paradigm and provide reliable signals for large-scale multi-task reinforcement learning.
- Controlled modality studies across more than 30 image, video, and interleaved generators show video generation excels at persistent spatiotemporal state tracking, while interleaved generation is more compute-efficient.
- All data, models, scorers, and code are publicly released.

## Full text

Native visual reasoning treats visual generation as the medium of reasoning itself: visual states (i.e. images and videos) are not merely inputs to be understood or outputs to be rendered, but first-class substrates for problem solving beyond language. Yet progress remains bottlenecked by the lack of scalable training tasks, reliable feedback, and controlled comparisons across generative substrates. In this work, we introduce VBVR-Pro, a closed-loop testbed that makes native visual reasoning through generation trainable, verifiable, optimizable, and experimentally controllable. 1) Task scaling. VBVR-Pro turns visual reasoning into a controlled task space of 300 procedurally generated tasks. Models trained on VBVR-Pro show strong transfer beyond the proposed suite across seven external visual reasoning benchmarks such as RISE-Video, MME-CoF-Pro, and BabyVision. 2) Verifiable rewards. VBVR-Pro provides verifiable reward scorers for task-grounded evaluation. Through a systematic study of leading MLLMs as judges, we identify recurring failure modes of the prevalent VLM-as-a-judge paradigm. In contrast, the proposed scorers are grounded in deterministic, task-specific rules, achieve fine-grained alignment with human judgments. Importantly, they serve as reliable reward signals for large-scale multi-task reinforcement learning and demonstrate stronger post-RL performance across visual reasoning tasks. 3) Mechanism study. VBVR-Pro enables controlled modality studies across more than 30 image, video, and interleaved generators. Our analysis shows that video generation remains strongest for tasks requiring persistent spatiotemporal state tracking, while interleaved generation provides a compute-efficient alternative. Critically, ablations and probing suggest the presence of vision-native trajectories that are crucial to visual reasoning. We release all data, models, scorers, and code.

---

Verify against the original before citing; curated by Astral Feed: https://feed.astralor.com/items/itm_615e5519323f532c262da8b4