GridAgent: A 2D Grid-Based Game Framework And Benchmark For Multimodal Large Language Models

Hengwei Ye, Yuanting Guan, Yuxuan Ge, Zheng Tian

Keywords: MLLM; benchmark; game

Abstract:

Multimodal Large Language Models (MLLMs) integrate the linguistic capabilities of LLMs with the ability to process multimodal data, enabling them to address a wider array of tasks. However, a comprehensive and standardized benchmark for evaluating MLLMs' complex visual reasoning performance in multimodal tasks has yet to be established. We introduce GridAgent, a versatile 2D grid-based framework that serves as a benchmark for assessing five essential capabilities of MLLMs: execution, perception reasoning, memory, learning, and planning. The framework includes twelve unique game tasks specifically designed to avoid overlap with the model's pre-training corpus. Each task targets at least one core competency and is enriched with diverse semantic information. Additionally, the game layouts are randomly generated, ensuring a more rigorous and authentic assessment of the MLLMs' capabilities. Experimental results indicate that although certain MLLMs excel in specific capabilities, none exhibit a comprehensive skill set comparable to the human baseline. Our work can be seen at: https://iclr2025gridagent.github.io/GridAgent-website.

Primary Area: datasets and benchmarks

Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.

Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide.

Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.

No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.

Submission Number: 11029