VisualSphinx: Large-Scale Synthetic Vision Logic Puzzles for RL

Yichen Feng; Zhangchen Xu; Fengqing Jiang; Yuetai Li; Bhaskar Ramasubramanian; Luyao Niu; Bill Yuchen Lin; Radha Poovendran

cs.CVMay 29, 2025

VisualSphinx: Large-Scale Synthetic Vision Logic Puzzles for RL

Authors: Yichen Feng, Zhangchen Xu, Fengqing Jiang, Yuetai Li, Bhaskar Ramasubramanian, Luyao Niu, Bill Yuchen Lin, Radha Poovendran

Organizations: University of Washington · Western Washington University

Abstract

Vision language models (VLMs) are expected to perform effective multimodal reasoning and make logically coherent decisions, which is critical to tasks such as diagram understanding and spatial problem solving. However, current VLM reasoning lacks large-scale and well-structured training datasets. To bridge this gap, we propose VisualSphinx, a first-of-its-kind large-scale synthetic visual logical reasoning training dataset. To tackle the challenge of image synthesis with grounded answers, we propose a rule-to-image synthesis pipeline that extracts and expands puzzle rules from seed questions and generates executable rendering programs for puzzle assembly. Using standard GRPO with binary correctness rewards as a validation recipe, experiments show that VLMs trained on VisualSphinx benefit from the logical coherence and readability of our dataset and exhibit improved performance on logical reasoning tasks. The enhanced reasoning capabilities developed from VisualSphinx also benefit other reasoning tasks such as algebraic reasoning, arithmetic reasoning and geometry reasoning.

Figures & tables

Figure 1: Example VisualSphinx instances within each reasoning category. Each visual logic puzzle comprises a text prompt, a graphical question stem with four images and a question mark, and four candidate choices of graphical answers.

Figure 2: This figure illustrates the four-stage pipeline for generating VisualSphinx . In Step 1, we collect 4K seed puzzles with explanations and abstract them into structured rule descriptions using LLMs. In Step 2, we apply a rule-level genetic algorithm to cross over, mutate and diversify the seed rules, scaling them to 40K high-quality rules. In Step 3, each rule is paired with a rendering style and used to generate five correct and three incorrect images via LLM-generated Python scripts. The fifth correct image is designated as the answer option, while the three rule-breaking images serve as distractors. After deduplication, we obtain 110K image groups. In Step 4, we assemble puzzles from each group using three complementary strategies: default assembly, shuffled answer variants, and expanded distractor sets. This results in over 660K visual logic puzzles, enabling robust and diverse training for multimodal reasoning models.

Figure 5: This figure demonstrates the statistics of readability and logical coherence.

Figure 6: This figure demonstrates the statistics of pass rates.

Panel of Table 1 — Table 1: Accuracy of vision models on VisualSphinx - Bench across difficulty levels. Qwen2.5-VL-3B-VisualSphinx and Qwen2.5-VL-7B-VisualSphinx represent our RL-trained models fine-tuned on VisualSphinx . The 7B model reaches GPT-4.1-comparable overall performance, performs strongest on Easy examples, and remains lower on Medium and Hard examples. Numbers with the best performance for each difficulty level are highlighted in bold , while those with the second-highest performance are underlined .

Vision – Math
Benchmark	Qwen2.5-VL-7B	Qwen2.5-VL-7B	Qwen2.5-VL-3B	Qwen2.5-VL-3B
	Instruct	VisualSphinx	Instruct	VisualSphinx
Mathverse vision	27.79	34.77	23.35	22.34
Mathvista mini	60.00	65.30	33.60	47.60
We-Math mini	61.21	62.30	54.20	59.20
Vision – Puzzle/Logic

Table 3: Performance on different benchmarks (accuracy, %). RAVEN reports the average over seven layouts; layout-level results are provided in Appendix Table 10 .

Vision – Math
Benchmark	Qwen2.5-VL-7B	SFT	SFT	Qwen2.5-VL-7B
	Instruct	Random	Reject	VisualSphinx
MathVerse vision	27.79	30.20	23.35	34.77
Mathvista mini	60.00	56.70	57.80	65.30
We-Math mini	61.21	51.90	44.48	62.30
Vision – Puzzle

Table 4: Ablation on VisualSphinx : SFT vs. RL across five benchmarks (accuracy, %).

Appendix figures & tables16 assets

Supplementary material from the paper’s appendix.

Appendix

Stage	Output Count	Description
Seed Questions	3,904	Raw visual logic questions
Filtered Seed Rules	2,398	After rewriting and verification
Generated Rules	60,339	After rule-level evolution
Retained Rules	41,287	After deduplication and scoring
Style 1 Image Groups	39,307	Successfully rendered
Style 2 Image Groups	38,918	Successfully rendered

Appendix

Table 5: Detailed statistics at each stage of the VisualSphinx pipeline.

Figure 12: Five different visual patterns of seeds

Visual Pattern	Reasoning Style	Count
Horizontal square	Deductive	417
Horizontal square	Inductive	447
Nine-square grid	Deductive	107
Nine-square grid	Inductive	199
Analogy	Deductive	143
Analogy	Inductive	118

Appendix

Table 6: Distribution of rule types based on visual pattern and reasoning style. Entries involving Others in either dimension are collapsed into the final row.

Figure 13: Rule distance.

Figure 14: Two puzzles that are visually similar but semantically different.

Figure 15: Example comparison of five models on the same rule prompt for image generation.

Hyper-parameter	Value
Learning Rate	$1\times 10^{-6}$
Evaluated Checkpoint (Step)	$256$
Extended 7B Run (Steps)	$320$
Number of Devices	$8$
Rollout Batch Size	$64$
PPO Mini Batch Size	$64$

Appendix

Table 7: This table shows the hyper-parameters for RL training.

Hyper-parameter	Value
Effective Batch Size	$128$ (= $8$ per device $\times$ $8$ GPUs $\times$ GA $2$ )
Per-Device Train Batch Size	$8$
Gradient Accumulation Steps	$2$
Learning Rate	$2\times 10^{-5}$
Epochs	$3$
LR Scheduler	cosine

Appendix

Table 8: Hyper-parameters for SFT training.

Model	Accuracy (%)	$\Delta$ vs. Instruct
Qwen2.5-VL-3B-Instruct	23.94	–
Qwen2.5-VL-3B-VisualSphinx	49.30	+25.36
Qwen2.5-VL-7B-Instruct	28.17	–
Qwen2.5-VL-7B-VisualSphinx	56.34	+28.17

Appendix

Table 9: Accuracy on the seed-rule-disjoint subset of VisualSphinx - Bench . The 71 examples in this subset have underlying seed rules that are entirely disjoint from the seed rules used to generate the 10K RL training set.

Layout	3B Instruct	3B- VisualSphinx	7B Instruct	7B- VisualSphinx
center_single	17.15	24.90	28.55	33.65
distribute_four	11.75	17.00	18.90	21.80
distribute_nine	11.40	13.30	15.55	16.90
in_center_single_out_center_single	14.65	20.90	26.30	36.55
in_distribute_four_out_center_single	13.20	22.15	23.95	33.55
left_center_single_right_center_single	11.40	12.65	13.85	22.20

Appendix

Table 10: Layout-level external visual-logic evaluation on RAVEN. Accuracy is reported in percentage points across seven RAVEN layouts. The average row is reported as the RAVEN result in Table 3 .

Figure 16: Detailed performance breakdown on MathVista-testmini before and after RL training on VisualSphinx . The radar chart shows improvements across multiple reasoning types, with particularly strong gains in logical reasoning (LOG), algebraic reasoning (ALG), and arithmetic reasoning (ARI).

Synthetic Data Generation for VLMs

See all 13 papers

Vero: An Open RL Recipe for General Visual Reasoning
Apr 6, 2026Gabriel Sarch, Linrong Cai, Qunzhong Wang +3Reasoning for VLMs Visual Reasoning
Princeton University
VisionFoundry: Teaching VLMs Visual Perception with Synthetic Images
Apr 10, 2026Guanyu Zhou, Yida Yin, Wenhao Chai +3Visual Question Answering Synthetic Data Augmentation
Princeton University · New York University
HopChain: Multi-Hop Data Synthesis for Generalizable Vision-Language Reasoning
Mar 17, 2026Shenzhi Wang, Shixuan Liu, Jing Zhou +8Reasoning for VLMs VLM Reasoning
Qwen Team, Alibaba Inc. · LeapLab, Tsinghua University

Apr 6, 2026J/K move · Enter open · S save

Vero: An Open RL Recipe for General Visual Reasoning

Gabriel Sarch, Linrong Cai, Qunzhong Wang, Haoyang Wu+2

Princeton University

Reasoning for VLMs Visual Reasoning Large Vision-Language Models VLM Reasoning RL for VLMs

Open paper AskPDF arXiv

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 2.9-5.4 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. All data, code, and models are publicly available.

Abstract

Synthetic Data Generation for VLMs

Explore similar work

Vero: An Open RL Recipe for General Visual Reasoning

VisionFoundry: Teaching VLMs Visual Perception with Synthetic Images

HopChain: Multi-Hop Data Synthesis for Generalizable Vision-Language Reasoning