Mexa Eval

mexa-eval

MEXA: Towards General Multimodal Reasoning with Dynamic Multi-Expert Aggregation — Yu et al. (2025) (arXiv:2506.17113, 2025)

What this evaluates

Evaluates a training-free, dynamic multi-expert aggregation framework for multimodal reasoning. It tests the system's ability to select specialized pre-trained experts and synthesize their outputs across video, audio, 3D, and medical domains without fine-tuning.

Datasets

• Video-MMMU — total ?; splits: test (-1)
• MMAU — total ?; splits: test (-1)
• SQA3D — total ?; splits: test (-1)
• M3D — total ?; splits: test (-1)

Metrics

• accuracy (primary) — range: percent
  • Percentage of correctly answered multiple-choice questions out of the total number of questions in the benchmark.

Input / output format

Input: Multimodal inputs (video, audio, 3D scenes, or medical scans) paired with multiple-choice questions.

Output: A single selected option from the provided multiple-choice answers.

Scoring recipe

def calculate_accuracy(predictions, gold_labels):
    correct = sum(1 for p, g in zip(predictions, gold_labels) if p == g)
    return (correct / len(gold_labels)) * 100

Common pitfalls

• The framework relies on external captioners for each modality; mismatched caption quality or prompt engineering can bottleneck performance regardless of the router/aggregator strength.
• Evaluation is strictly multiple-choice; open-ended generation or free-form reasoning capabilities are not measured.
• Performance is highly sensitive to the choice of router and aggregator (e.g., GPT-4o vs Qwen2.5-VL, DeepSeek vs GPT-4o); swapping these without re-evaluation may yield significantly different results.

Evidence (verbatim from paper)

We evaluate MEXA on all datasets under the multiple-choice QA setting, and report performance based on standard accuracy metrics across all experiments.

Citation

@misc{yu2025mexa,
  title={MEXA: Towards General Multimodal Reasoning with Dynamic Multi-Expert Aggregation},
  author={Yu et al. (2025)},
  year={2025},
  note={arXiv:2506.17113}
}

• arXiv: 2506.17113