license: apache-2.0
annotations_creators:
- expert-generated
- found
language_creators:
- expert-generated
- found
task_categories:
- question-answering
- multiple-choice
- visual-question-answering
language:
- en
- ch
tags:
- think-with-image
- agentic reasoning
- reasoning
- multi-modal-qa
pretty_name: TIR-Bench
size_categories:
- 1K<n<10K
configs:
- config_name: default
data_files:
- split: test
path: TIR-Bench.json
TIR-Bench: A Comprehensive Benchmark for Agentic Thinking-with-Images Reasoning
Introduction:
TIR-Bench is a comprehensive benchmark designed to evaluate the "thinking-with-images" capabilities of Multimodal Large Language Models (MLLMs), addressing a gap left by existing benchmarks like Visual Search which only test basic operations. As models like OpenAI o3 begin to intelligently create and operate tools to transform images for problem-solving, TIR-Bench provides 13 diverse tasks that each require novel tool use for image processing and manipulation within a chain-of-thought. Our evaluation of 22 leading MLLMs (including open-sourced, proprietary, and tool-augmented models) shows that TIR-Bench is universally challenging and that strong performance requires genuine agentic thinking-with-images capabilities. This repository contains the full benchmark, evaluation scripts, and a pilot study comparing direct versus agentic fine-tuning for this advanced reasoning.
Paper Link: https://arxiv.org/abs/2511.01833
If you use this benchmark in your research, please consider citing it as follows:
@article{li2025tir,
title={TIR-Bench: A Comprehensive Benchmark for Agentic Thinking-with-Images Reasoning},
author={Li, Ming and Zhong, Jike and Zhao, Shitian and Zhang, Haoquan and Lin, Shaoheng and Lai, Yuxiang and Chen, Wei and Psounis, Konstantinos and Zhang, Kaipeng},
journal={arXiv preprint arXiv:2511.01833},
year={2025}
}