MMSearch-Plus✨: Benchmarking Provenance-Aware Search for Multimodal Browsing Agents

Official repository for the paper "MMSearch-Plus: Benchmarking Provenance-Aware Search for Multimodal Browsing Agents".

🌟 For more details, please refer to the project page with examples: https://mmsearch-plus.github.io/.

[🌐 Webpage] [📖 Paper] [🤗 Huggingface Dataset] [🏆 Leaderboard]

💥 News

• [2025.09.26] 🔥 We update the arXiv paper and release all MMSearch-Plus data samples in huggingface dataset.
• [2025.08.29] 🚀 We release the arXiv paper.

📌 ToDo

• Agentic rollout framework code
• Evaluation script
• Set-of-Mark annotations

Dataset corrections — 2026-10-17

This revision corrects two answer annotations reported in GitHub issue #4:

Published sample	Zero-based row	Previous answer	Corrected answer
Q24	23	Zhu Yi, Liu Jikang	2021
Q247	246	Empty answer list	Approximately 1.4 g/cm³

Q24's source annotation contains a year question and year-based scoring criteria for 2021, but retained the player names from an earlier question. The correction uses those source scoring criteria supplied by the dataset collaborator.

Q247 corresponds to raw entry_index=286, not raw entry 247. Its source is the Sandatlas sand comparison: sample 6 is from Utah's Coral Pink Sand Dunes and sample 8 is garnet sand from Emerald Creek, Idaho. The collaborator identified the orange and burgundy minerals as quartz and garnet, respectively, and supplied densities of approximately 2.65 and 4.0–4.1 g/cm³. Their absolute difference is 1.35–1.45 g/cm³, rounded to approximately 1.4 g/cm³. These are mineral densities, not bulk sand densities. The linked comparison establishes the sample identities; the density values are from the collaborator's correction notes. Evaluation should accept this range and equivalent units.

Only these two answer fields changed; the dataset still contains 311 rows in the same order, with the same questions, images, and other metadata. Answers retain the existing encryption format. For reproducing results on the original annotations, pin revision 161a60d469b67cd14a0621ae2ab66d902df02df9. The state.json disk-loading manifest was also corrected to list the five published Arrow shards; it previously referenced two nonexistent files. Report the dataset revision with evaluation results; historical results have not been recomputed as part of this annotation correction.

Usage

⚠️ Important: This dataset is encrypted to prevent data contamination. However, decryption is handled transparently by the dataset loader.

Dataset Usage

For better compatibility with newer versions of the datasets library, we provide explicit decryption functions, downloadable from our GitHub/HF repo.

wget https://raw.githubusercontent.com/mmsearch-plus/MMSearch-Plus/main/decrypt_after_load.py

import os
from datasets import load_dataset
from decrypt_after_load import decrypt_dataset

encrypted_dataset = load_dataset("Cie1/MMSearch-Plus", split='train')
decrypted_dataset = decrypt_dataset(
    encrypted_dataset=encrypted_dataset,
    canary='your_canary_string' # Set the canary string (hint: it's the name of this repo without username)
)

# Access a sample
sample = decrypted_dataset[0]
print(f"Question: {sample['question']}")
print(f"Answer: {sample['answer']}")
print(f"Category: {sample['category']}")
print(f"Number of images: {sample['num_images']}")

# Access images (PIL Image objects)
sample['img_1'].show()  # Display the first image

👀 About MMSearch-Plus

MMSearch-Plus is a challenging benchmark designed to test multimodal browsing agents' ability to perform genuine visual reasoning. Unlike existing benchmarks where many tasks can be solved with text-only approaches, MMSearch-Plus requires models to extract and use fine-grained visual cues through iterative image-text retrieval.

Key Features

🔍 Genuine Multimodal Reasoning: 311 carefully curated tasks that cannot be solved without visual understanding

🎯 Fine-grained Visual Analysis: Questions require extracting spatial cues and temporal traces from images to find out-of-image facts like events, dates, and venues

🛠️ Agent Framework: Model-agnostic web agent with standard browsing tools (text search, image search, zoom-in)

📍 Set-of-Mark (SoM) Module: Enables provenance-aware cropping and targeted searches with human-verified bounding box annotations

Dataset Structure

Each sample contains:

• Quuestion text and images
• Ground truth answers and alternative valid responses
• Metadata including arXiv id (if an event is a paper), video URL (if an event is a video), area and subfield

Performance Results

Evaluation of closed- and open-source MLLMs shows:

• Best accuracy is achieved by o3 with full rollout: 36.0% (indicating significant room for improvement)
• SoM integration provides consistent gains up to +3.9 points
• Models struggle with multi-step visual reasoning and cross-modal information integration

The overview of three paradigms for multimodal browsing tasks that demand fine-grained visual reasoning.

The overview of an example trajectory for a task in MMSearch-Plus.

🏆 Leaderboard

Contributing to the Leaderboard

🚨 The Leaderboard is continuously being updated, welcoming the contribution of your excellent LMMs!

🔖 Citation

If you find MMSearch-Plus useful for your research and applications, please kindly cite using this BibTeX:

@article{tao2025mmsearch,
  title={MMSearch-Plus: A Simple Yet Challenging Benchmark for Multimodal Browsing Agents},
  author={Tao, Xijia and Teng, Yihua and Su, Xinxing and Fu, Xinyu and Wu, Jihao and Tao, Chaofan and Liu, Ziru and Bai, Haoli and Liu, Rui and Kong, Lingpeng},
  journal={arXiv preprint arXiv:2508.21475},
  year={2025}
}