visual-reasoning-benchmark-results
收藏资源简介:
Visual Reasoning Benchmark Suite v3.3 是一个综合性视觉推理基准测试数据集,包含2005个任务,涵盖12个等权重的Track。数据集旨在系统评估模型在多种视觉推理任务上的能力。数据内容包括12类任务:图形补全(394题)、空间生成(56题)、迷宫(分初级、中级、高级,各64题)、数独推理(78题)、数织推理(150题,含45简单/60中等/45困难)、七巧板推理(150题,含45简单/60中等/45困难)、棋盘游戏推理(300题)、火柴棒推理(300题)、正交投影推理(90题)以及数学视觉推理(295题)。其中,数织和七巧板是v3.3版本新增的任务,七巧板题目的生成经过严格流程,确保唯一性和复杂度。数据以文件形式组织,包含题目图、答案图、掩码和几何信息等。评估体系采用12类等权重评分,主排名使用“Macro Overall”分数,同时提供“Micro Overall”作为辅助指标,具体任务有定制评估方法。数据集配套完整代码工具链,支持使用视觉模型进行题目生成、结果评估和性能报告,主要用于评测和比较不同模型在复杂视觉推理任务上的综合性能。
Visual Reasoning Benchmark Suite v3.3 is a comprehensive visual reasoning benchmark dataset containing 2005 tasks spanning 12 equally-weighted tracks. It aims to systematically evaluate the capabilities of models across diverse visual reasoning tasks. The dataset includes 12 categories of tasks: 1. Shape Completion (394 tasks) 2. Spatial Generation (56 tasks) 3. Maze (divided into primary, intermediate and advanced levels, with 64 tasks per level) 4. Sudoku Reasoning (78 tasks) 5. Nonogram Reasoning (150 tasks, including 45 easy, 60 medium and 45 hard instances) 6. Tangram Reasoning (150 tasks, including 45 easy, 60 medium and 45 hard instances) 7. Board Game Reasoning (300 tasks) 8. Matchstick Puzzle Reasoning (300 tasks) 9. Orthographic Projection Reasoning (90 tasks) 10. Mathematical Visual Reasoning (295 tasks) Among them, Nonogram and Tangram are newly added tasks in version v3.3. The generation of Tangram tasks follows a strict workflow to ensure their uniqueness and complexity. The data is organized in file format, containing problem images, answer images, masks, geometric information and other relevant contents. The evaluation system adopts equally-weighted scoring for the 12 categories. The "Macro Overall" score is used for the main ranking, while "Micro Overall" is provided as a supplementary metric, with customized evaluation methods for specific tasks. The dataset is accompanied by a complete code toolchain that supports task generation, result evaluation and performance reporting using visual models. It is primarily designed to evaluate and compare the comprehensive performance of different models on complex visual reasoning tasks.
数据集概述
Visual Reasoning Benchmark Suite v3.3 是一个多任务视觉推理基准测试集,包含 2005 个任务,覆盖 12 个 Track。该版本以用户最新上传的 visual_reasoning_benchmark_suite_v3_修改 为基础,保持原有 3283 个数据文件字节级不变,并新增了 Nonogram(数织)和 Tangram(七巧板)两类任务。
任务与数量
| 任务名称 | 数量 |
|---|---|
| figure_completion | 394 |
| spatial_generation | 56 |
| maze_beginner | 64 |
| maze_intermediate | 64 |
| maze_advanced | 64 |
| sudoku_reasoning | 78 |
| nonogram_reasoning | 150 |
| tangram_reasoning | 150 |
| board_game_reasoning | 300 |
| matchstick_reasoning | 300 |
| orthographic_reasoning | 90 |
| math_visual_reasoning | 295 |
| 总计 | 2005 |
评分规则
- 正式主排名:Macro Overall,计算公式为 12 个 Track 平均分之和除以 12。每个 Track 权重相同(各占约 8.33%),题量不影响权重。
- 保留指标:Micro Overall,仅作为诊断参考,不用于正式排名。
目录结构
visual_reasoning_benchmark_suite_v3_3_equal_weight_2005/ ├── datasets/ │ ├── civil_service/ │ ├── maze/ │ ├── sudoku/ │ ├── nonogram/ │ ├── tangram/ │ ├── board_game/ │ ├── matchsticks/ │ ├── orthographic/ │ └── mathematical_proof/ ├── code/ │ ├── benchmark.py │ ├── generate_all.py │ ├── evaluate_all.py │ ├── validate_all.py │ ├── report_all.py │ ├── tools/ │ │ ├── build_nonogram_benchmark.py │ │ └── build_tangram_benchmark.py │ ├── evaluators/ │ │ ├── nonogram.py │ │ └── tangram.py │ └── report/ └── validation_summary.json
安装与使用
- 环境要求:Python 3.10+。
- 基本命令:
- 查看全部任务:
python benchmark.py list - 验证全部任务:
python benchmark.py validate --tasks all - 生成奖励任务(Nonogram/Tangram):使用
python benchmark.py build-nonogram或python benchmark.py build-tangram,支持自定义随机种子和候选池倍数。 - 生成模型输出:指定 provider、model、run-name,支持按任务限制数量(
--limit-per-task)和并发数(--workers),支持断点续跑。 - 评估结果:
python benchmark.py evaluate,指定 judge-model、passes 和 workers。 - 生成报告:
python benchmark.py report,指定 runs 和 judge-model,输出排行榜、雷达图等。
- 查看全部任务:
评估方法
- Nonogram 评估:优先程序化读取黑白格,不确定时回退 GPT-5.5 网格转录,并由 Python 验证行列约束。主要指标包括 Exact Solve Rate、Cell Accuracy、Row Constraint Accuracy、Column Constraint Accuracy。
- Tangram 评估:混合评估流程:
- 计算机视觉计算目标覆盖率、IoU、越界、下方残留和颜色拼板数量。
- 明确的满分/零分样本直接程序判定。
- 边界样本回退 GPT-5.5,判断是否使用全部七块、保持形状、无重叠、无越界并完整填满目标。
- 非唯一解,其他合法拼法也可满分。
输出文件
报告生成后输出:
leaderboard.mdleaderboard.csvraw_scores.csvscore_audit.jsonoverall_bar.pngradar_chart.png




