wellingyuan/SSI-Bench

Thinking in Structures: Evaluating Spatial Intelligence through Reasoning on Constrained Manifolds

SSI-Bench is constructed from complex real-world 3D structures, where feasible configurations are tightly governed by geometric, topological, and physical constraints.

• Project page: https://ssi-bench.github.io/
• Code & evaluation: https://github.com/ccyydd/SSI-Bench
• Paper: https://arxiv.org/abs/2602.07864

News

• 🔥 [2026-2-10]: We released our paper, benchmark, and evaluation codes.

Load Dataset

from datasets import load_dataset

dataset = load_dataset("cyang203912/SSI-Bench")
print(dataset)

After downloading the parquet file, read each record, decode images from binary, and save them as JPG files.

import pandas as pd
import os

df = pd.read_parquet("SSI_Bench.parquet")

output_dir = "./images"
os.makedirs(output_dir, exist_ok=True)

for _, row in df.iterrows():
    index_val = row["index"]
    images = row["image"]
    question = row["question"]
    answer = row["answer"]
    annotation_color = row["annotation_color"]
    category = row["category"]
    task = row["task"]

    image_paths = []
    if images is not None:
        for n, img_data in enumerate(images):
            image_path = f"{output_dir}/{index_val}_{n}.jpg"
            with open(image_path, "wb") as f:
                f.write(img_data)
            image_paths.append(image_path)
    else:
        image_paths = []

    print(f"index: {index_val}")
    print(f"image: {image_paths}")
    print(f"question: {question}")
    print(f"answer: {answer}")
    print(f"annotation_color: {annotation_color}")
    print(f"category: {category}")
    print(f"task: {task}")
    print("-" * 50)

Usage

To evaluate, follow the scripts in the code repository: https://github.com/ccyydd/SSI-Bench.

Leaderboard

Model	Avg. (%)	Type
Human Performance	91.60	Baseline
Gemini-3-Flash	33.60	Proprietary
Gemini-3-Pro	29.50	Proprietary
GPT-5.2	29.10	Proprietary
Gemini-2.5-Pro	26.10	Proprietary
GPT-5 mini	25.90	Proprietary
Seed-1.8	25.90	Proprietary
GPT-4o	22.60	Proprietary
GPT-4.1	22.40	Proprietary
Gemini-2.5-Flash	22.30	Proprietary
GLM-4.6V	22.20	Open-source
Qwen3-VL-235B-A22B	21.90	Open-source
GLM-4.5V	21.40	Open-source
GLM-4.6V-Flash	21.10	Open-source
Qwen3-VL-4B	20.70	Open-source
InternVL3.5-30B-A3B	20.70	Open-source
Qwen3-VL-30B-A3B	20.60	Open-source
Llama-4-Scout-17B-16E	20.60	Open-source
Gemma-3-27B	20.50	Open-source
InternVL3.5-8B	20.20	Open-source
Claude-Sonnet-4.5	19.90	Proprietary
Gemma-3-4B	19.70	Open-source
Qwen3-VL-8B	19.20	Open-source
Qwen3-VL-2B	19.20	Open-source
InternVL3.5-38B	19.00	Open-source
InternVL3.5-241B-A28B	18.30	Open-source
InternVL3.5-14B	17.90	Open-source
Gemma-3-12B	17.30	Open-source
LLaVA-Onevision-72B	17.20	Open-source
InternVL3.5-4B	16.80	Open-source
LLaVA-Onevision-7B	16.50	Open-source
Random Guessing	12.85	Baseline
InternVL3.5-2B	11.10	Open-source

Citation

@article{yang2026thinking,
  title={Thinking in Structures: Evaluating Spatial Intelligence through Reasoning on Constrained Manifolds},
  author={Chen Yang and Guanxin Lin and Youquan He and Peiyao Chen and Guanghe Liu and Yufan Mo and Zhouyuan Xu and Linhao Wang and Guohui Zhang and Zihang Zhang and Shenxiang Zeng and Chen Wang and Jiansheng Fan},
  journal={arXiv preprint arXiv:2602.07864},
  year={2026}
}