mmJEE-Eval: A Bilingual Multimodal Benchmark for Exam-Style Evaluation of Vision-Language Models

Paper: mmJEE-Eval: A Bilingual Multimodal Benchmark for Evaluating Scientific Reasoning in Vision-Language Models Code: https://github.com/ArkaMukherjee0/mmJEE-Eval Project Page: https://mmjee-eval.github.io

Introduction

mmJEE-Eval is a multimodal and bilingual dataset for LLM evaluation comprising 1,460 challenging questions from seven years (2019-2025) of India's JEE Advanced competitive examination. We evaluate 17 state-of-the-art VLMs, finding that open models (from 7B-400B) struggle significantly (maxing at 40-50%) as compared to frontier models from Google and OpenAI (77-84%). mmJEE-Eval is significantly more challenging than the text-only JEEBench, the only other well-established dataset on JEE Advanced problems, with performance drops of 18-56% across all models. Our findings, especially metacognitive self-correction abilities, cross-lingual consistency, and human evaluation of reasoning quality, demonstrate that contemporary VLMs still show authentic scientific reasoning deficits despite strong question-solving capabilities (as evidenced by high Pass@K accuracies), establishing mmJEE-Eval as a challenging complementary benchmark that effectively discriminates between model capabilities.

Sample Usage

You can load the dataset using the Hugging Face datasets library:

from datasets import load_dataset

# Load the entire dataset
dataset = load_dataset("ArkaMukherjee/mmJEE-Eval")

# Access the training split
train_dataset = dataset["train"]

# Print an example
print(train_dataset[0])

# To run evaluation scripts, please refer to the official GitHub repository:
# https://github.com/ArkaMukherjee0/mmJEE-Eval