🧠 ReasonBench: Benchmarking and Improving Visual Language Models for Complex Graphic Reasoning

background

image：background

🌐 Overview

ReasonBench is a comprehensive benchmark designed to evaluate Visual Language Models (VLMs) on complex graphical reasoning tasks. It contains 1,613 problems collected from real-world intelligence tests, covering 11 core cognitive dimensions and 29 task types. This benchmark provides a robust framework for assessing VLMs' spatial, relational, and abstract reasoning capabilities.

Dataset Type: Visual Language Reasoning · Graphical Reasoning · Benchmark Evaluation

Paper Link:https://arxiv.org/abs/2508.00323

📊 Dataset Structure

Core Cognitive Dimensions & Task Types

Cognitive Dimension	Task Type	Count
Positional Patterns	Translation	94
	Rotation	56
	Combination	30
Stylistic Patterns	Crossing	54
	Addition/Subtraction	67
	Black/White Operation	63
Attribute Patterns	Symmetry	109
	Open/Close State	19
	Combination	6
Quantitative Patterns	Lines	173
	Faces	137
	Points	66
	Elements	94
	Combination	50
Spatial Patterns	Cubes	109
	3D	46
	Polyhedrons	17
	Three Views	40
	Cross-Sections	35
	Spatial Quantitative Trans.	10
Special Patterns	2D Combination	31
	Figure Relations	40
Alphanumeric	Alphanumeric	27
B&W Blocks	Black & White Blocks	32
Other Patterns	Comprehensive	34
MENSA	Task 1	35
	Task 2	39
Raven	Task 1	40
	Task 2	60

🖼️ Input Formats

Format	Description
Integrated Format	Presents questions and options in a single image for holistic processing
Separated Format	Splits questions and options into multiple images for step-by-step reasoning

🔍 Key Features

Multi-format Evaluation: Supports both integrated and separated input formats
Full Accessibility: Provides public URLs for all images (questions, options, and combined sets)
Human Baseline: Includes human performance metrics for comparison
Diverse Tasks: Covers 29 distinct reasoning task types across 11 cognitive dimensions

🚀 Usage（GPT-4o example）

import base64
import requests
import os
from openai import OpenAI  # Requires openai>=1.0.0

# Configuration
api_key = os.getenv("OPENAI_API_KEY")
if not api_key:
    raise ValueError("Missing OPENAI_API_KEY environment variable")

# Initialize client (official SDK approach)
client = OpenAI(api_key=api_key)

def process_image_question(image_path: str, question: str, max_tokens=300):
    """Send image and question to GPT-4o API"""
    # Encode image to base64
    base64_image = base64.b64encode(open(image_path, "rb").read()).decode("utf-8")

    # Construct messages payload
    messages = [
        {
            "role": "user",
            "content": [
                {"type": "text", "text": question},
                {
                    "type": "image_url",
                    "image_url": {
                        "url": f"data:image/jpeg;base64,{base64_image}",
                        "detail": "auto"  # Options: low, high, auto
                    }
                }
            ]
        }
    ]

    # Make API request
    response = client.chat.completions.create(
        model="gpt-4o",
        messages=messages,
        max_tokens=max_tokens
    )
    
    return response.choices[0].message.content

# Example usage
if __name__ == "__main__":
    image_path = "path/to/your/image.jpg"  # Update with actual path
    user_question = "What's in this image?"  # Customize your question
    
    try:
        answer = process_image_question(image_path, user_question)
        print("AI Response:", answer)
    except Exception as e:
        print(f"Error: {str(e)}")

🧠 ReasonBench：复杂图形推理的视觉语言模型评估基准

🌐 概述

ReasonBench 是一个用于评估视觉语言模型（VLMs）在复杂图形推理任务表现的基准测试。数据集包含从真实智力测试中收集的 1,613个问题，覆盖11个核心认知维度和29种任务类型，为评估VLMs的空间、关系和抽象推理能力提供综合框架。

数据集类型：视觉语言推理 · 图形推理 · 基准评估

论文地址:https://arxiv.org/abs/2508.00323

📊 数据结构

核心认知维度与任务类型

认知维度	任务类型	数量
位置规律	平移	94
	旋转	56
	组合	30
样式规律	穿越	54
	加减法	67
	黑白运算	63
属性规律	对称	109
	开闭状态	19
	组合	6
数量规律	线	173
	面	137
	点	66
	元素	94
	组合	50
空间规律	立方体	109
	3D	46
	多面体	17
	三视图	40
	剖视图	35
	空间数量变换	10
特殊规律	2D组合	31
	图形关系	40
字母数字	字母数字	27
黑白块	黑白块	32
其他规律	综合	34
门萨	任务1	35
	任务2	39
瑞文	任务1	40
	任务2	60

🖼️ 输入格式

格式	描述
集成格式	问题与选项呈现在单个图形中，便于模型整体处理
分离格式	将问题与选项拆分为多个图形，测试分步推理能力

🔍 核心特性

多格式评估：支持整体式和分隔式两种输入格式
完全开放：公开所有格式的图片URL（题目、选项、题目+选项）
人类基准：提供人类准确率作为参考基准
多样化任务：覆盖11个认知维度的29种推理任务

🚀 使用示例(以openai GPT-4o为例)

import base64
import requests
import os

# 配置OpenAI API密钥
api_key = os.getenv("OPENAI_API_KEY")  # 建议将密钥存储在环境变量中
if not api_key:
    raise ValueError("请设置OPENAI_API_KEY环境变量")

# 图像处理函数
def encode_image(image_path):
    """将本地图像编码为base64字符串"""
    with open(image_path, "rb") as image_file:
        return base64.b64encode(image_file.read()).decode('utf-8')

# 示例图像路径和问题
image_path = "path/to/your/image.jpg"  # 替换为你的图像路径
question = "描述这张图片的内容"  # 替换为你的问题

# 构建API请求
headers = {
    "Content-Type": "application/json",
    "Authorization": f"Bearer {api_key}"
}

payload = {
    "model": "gpt-4o",  # 使用支持图像的模型
    "messages": [
        {
            "role": "user",
            "content": [
                {
                    "type": "text",
                    "text": question
                },
                {
                    "type": "image_url",
                    "image_url": {
                        "url": f"data:image/jpeg;base64,{encode_image(image_path)}"
                    }
                }
            ]
        }
    ],
    "max_tokens": 300  # 控制响应长度
}

# 发送请求
response = requests.post(
    "https://api.openai.com/v1/chat/completions",
    headers=headers,
    json=payload
)

# 处理响应
if response.status_code == 200:
    result = response.json()
    answer = result['choices'][0]['message']['content']
    print("AI回答:", answer)
else:
    print("请求失败，状态码:", response.status_code)
    print("错误信息:", response.text)

如果需要引用，请引用下列内容

{
  author        = {Jianyi Zhang and Xu Ji and Ziyin Zhou and Yuchen Zhou and Shubo Shi and Haoyu Wu and Zhen Li and Shizhao Liu},
  title         = {Oedipus and the Sphinx: Benchmarking and Improving Visual Language Models for Complex Graphic Reasoning},
  howpublished  = {arXiv preprint},
  archivePrefix = {arXiv},
  eprint        = {2508.00323},
  primaryClass  = {cs.AI},
  year          = {2025},
  note          = {arXiv:2508.00323v1 [cs.AI]},
  url           = {https://arxiv.org/abs/2508.00323}
}

pdf pdf