VBVR-Pro: A Scalable and Verifiable Suite for Native Visual Reasoning
Junxiang Xu, Ruisi Wang, Fanyi Pu, Maijunxian Wang, Ran Ji, Tongxi Zhou, Chenyang Gu, Jing Zuo, Hongcan Xiao, Yimeng Geng, Wanqi Yin, Wei Chen, Oscar Qian, Zhengan Yan, Ziqi Huang, Haiwen Diao, Liang Pan, Bo Li, Xiangyu Fan, Dezhi Luo, Fengyuan Yu, Zehong Zhao, Qingying Gao, Tinghui Zhu, Yilan Zhang, Jingqi Tong, Pinyuan Feng, Zhengze Jiang, Letian Wang, Ziyu Guo, Renrui Zhang, Jieneng Chen, Sonia Joseph, Constantin Venhoff, Saman Motamed, Mengyue Yang, Chandra Sripada, Alan Yuille, Philip Torr, Lvmin Zhang, Vikash Kumar, Daniel Khashabi, Nikolaus Kriegeskorte, Raphaël Millière, Vincent C. Müller, Anyi Rao, Quan Wang, Ziwei Liu, Dahua Lin, Lei Yang, Hokin Deng, Zhongang Cai
cs.CV, cs.AI, cs.LG, cs.MM, cs.RO
2026-08-27
VBVR-Pro trains native visual reasoning on 300 verifiable tasks; fine-tuned Wan2.2-I2V-A14B gains up to 28 points on seven external benchmarks.
Language models made token chains the default medium of reasoning. Spatial transforms, object persistence, and temporal tracking can be described in words. They are not native to words. Native visual reasoning treats generation as the medium: the model solves a problem by constructing and updating images or videos, and those visual states are working memory, not just inputs or a final render.
Three pieces of infrastructure were missing. Most visual-reasoning benchmarks are eval-only. VLM-as-a-judge looks convenient for open-ended outputs, then fails on exact counts, fine spatial relations, and rule checks, and is unstable even at temperature 0. Image, interleaved, and video generators are rarely trained and scored on one shared task distribution. VBVR-Pro is a closed-loop testbed for those three gaps.
Tasks are programs. Each of 300 generators samples grid size, object count, layout, appearance, and difficulty, then a solver writes the ground-truth answer. 150 tasks are reworked from the earlier VBVR suite; 150 are new. The new set is denser and deeper: median 80 connected color regions per sampled frame versus 12, 47% requiring multi-step search versus 7%, and 113 of 150 blinded pairs judged deeper. The same instance is rendered as video, as keyframe images, and as interleaved text, so the three generator families see one problem. Training uses 250 tasks times 5,000 instances, 1.25 million samples, plus a 50k RL split. VBVR-Pro-Bench holds 50 in-domain task families and 50 held-out families, 100 tasks, five instances each.
Scorers read semantic entities, not pixels. HSV segmentation, contours, OCR, and tracking recover color, shape, position, and count. Soft tasks take a weighted sum; hard constraints multiply, so a wall collision nearly zeros the score. VLM judges change scores on 54.6% to 92.8% of the same videos across reruns. The rule scorers do not.
RL needs semantic exploration: left or right in a maze. Low noise often changes texture and keeps the decision. Flow-SDE at high stochasticity wrecks the frame first. CPS (Coefficients-Preserving Sampling) splits the scheduler's noise coefficient into predicted noise plus fresh noise, so exploration rises without extra noise mass. The controlled comparison uses the same algorithm, data, and init; only the reward source switches from Qwen3.6-27B to the rule scorer.
Open-source generators are weak on the suite. Task-relevant data is the bottleneck. Fine-tuning nine of them for one epoch raises overall score by 0.290 on average, +0.401 in-domain and +0.179 out-of-domain. The strongest video model, VBVR-Pro-Wan2.2-I2V-A14B, reaches 0.670 overall from a 0.182 base. The strongest interleaved model, VBVR-Pro-SenseNova-U1, reaches 0.638. Among proprietary systems, Seedream-5.0-Pro scores 0.557 (image), Nano Banana Pro 0.564 (interleaved), Seedance 2.0 0.499 (video). Single-image generation lags most on transformation tasks: an endpoint image cannot hold the intermediate states.
An interleaved ablation makes the accounting clear. Compressing the visual trajectory to one image drops SenseNova from 0.638 to 0.527. Replacing intermediate text with placeholders barely moves the number, 0.629. At inference, deleting the intermediate image drops the score to 0.099; deleting or contradicting intermediate text leaves about 0.53. Removing the input image leaves 0.064. Visual trajectories are the working memory. Language is secondary.
Transfer on the strongest video model:
| Benchmark | Wan2.2 base | After VBVR-Pro | Veo 3.1 |
| V-ReasonBench | 10.21 | 38.22 | 24.25 |
| VideoThinkBench mini | 25.71 | 52.86 | 27.69 |
| MME-CoF-Pro | 24.76 | 48.39 | 55.90 |
| RULER-Bench | 57.89 | 66.96 | 65.19 |
| RISE-Video | 62.83 | 66.18 | 76.40 |
| BabyVision-Gen | 0.36 | 12.50 | – |
Shape analogies, block ordering, and robot pick-and-place look nothing like the synthetic puzzles. CLIP and DINOv2 nearest neighbors in the training set do not match the queries. The same backbone trained on the earlier VBVR suite scored 18.05 on V-ReasonBench; Pro training reaches 38.22. The extra 150 harder tasks are not more copies of the same generators.
RLVR from the already-SFT Wan2.2-TI2V-5B reaches 0.548 overall, against 0.503 for more SFT on the same 50k and 0.508 for VLM rewards. Out-of-domain moves from 0.328 to 0.377. One maze case goes from 0.2249 to 0.8692, with connected corridors in between.
Against humans, the rule scorer exceeds 0.60 per-vote agreement, versus 0.54 for GPT-5.5 and 0.52 for Gemini-3.1-Pro, about 78% of the 0.77 human ceiling. Model-level Spearman ρ is 0.95 in-domain and 1.00 out-of-domain.
People chasing generation-as-reasoning needed a substrate that can be trained, scored, and compared. Task coverage is a growing program. Rewards leave expensive, unstable VLMs for cheap deterministic rules. Multi-task RL actually runs on 50 tasks. If you adopt one piece, take the data and the scorers. Do not plug a commercial VLM in as the reward. Interleaved models approach video in-domain at lower cost. Video still wins on fine state tracking and out-of-domain transfer.
This is an infrastructure paper. The gains come from coverage and verifiable feedback, not a new backbone.
There is no dedicated limitations section. The numbers leave gaps. The best trained model is still far from human. Out-of-domain lags in-domain (0.532 vs 0.808 for the best video model). Scorers are tied to procedural renders and classical CV; they may not grade photoreal outputs. Interleaved text was written by Gemini-3.1-Pro, so text quality is capped there. RL covers only 50 in-domain tasks; OOD RL gains are a side effect. RISE-Video only moves from 62.83 to 66.18; the abstract's "often more than 20 points" describes the large lifts, not the seven-benchmark mean. Open models generally get one LoRA epoch; proprietary models are not fine-tuned on the same data, so Table 4 is not a ceiling comparison.