[Benchmark] Support MMReason

[HJYao00]

This PR adds evaluation support for the MMReason benchmark, accepted at ICCV 2025, to assess the reasoning capabilities of MLLMs.

Before submitting, I have tested that the code successfully run on Qwen2.5-VL and Qwen3-VL. The command to run the evaluation is as follows:

python3 run.py --data MMReason_testmini --model Qwen2.5-VL-7B-Instruct --verbose

python3 run.py --data MMReason_testmini --model Qwen3-VL-8B-Instruct --verbose

[FangXinyu-0913]

Hi @HJYao00, Have you evaluated popular MLLMs like Qwen-VL3 and InternVL3.5 on MMReason? If so, could you share the results from the VLMEvalKit along with your own evaluation data here? This would be very helpful for reproduction purposes.

[HJYao00]

Hi @FangXinyu-0913. I have evaluated the popular MLLMs (Qwen-VL-2.5, Qwen-VL-3, and InternVL-3.5) on the MMReason benchmark using VLMEvalKit. The performance of Qwen2.5-VL in VLMEvalKit almost matches the results reported in the paper. Thank you!

Model	Source of Results	Accuracy
Qwen-VL-2.5-7B	VLMEvalKit	16.9
Qwen-VL-2.5-7B	Paper	17.3
Qwen-VL-3-8B	VLMEvalKit	30.1
InternVL3.5-8B	VLMEvalKit	20.8