Multimodal Vqa Eval

multimodal-vqa-eval

Train a Unified Multimodal Data Quality Classifier with Synthetic Data — Wang et al. (2025) (arXiv:2510.15162, 2025)

What this evaluates

Evaluates the zero-shot and few-shot visual question answering capabilities of multimodal large language models (MLLMs). It probes scene and spatial understanding, OCR capabilities, commonsense knowledge reasoning, and multimodal in-context learning across diverse benchmarks.

Datasets

• GQA — total ?; splits: test (-1)
• VQA-v2 — total ?; splits: test (-1)
• VizWiz — total ?; splits: test (-1)
• TextVQA — total ?; splits: test (-1)
• OKVQA — total ?; splits: test (-1)
• POPE — total ?; splits: test (-1)
• MMMU (Val) — total ?; splits: val (-1)
• MMBench (Dev) — total ?; splits: dev (-1)
• MMStar — total ?; splits: test (-1)

Metrics

• accuracy (primary) — range: percent
  • Percentage of correctly answered questions. Computed as the mean score over 5 random seeds for demonstration example sampling in few-shot settings.

Input / output format

Input: Image(s) and text prompt (question or few-shot demonstrations + question).

Output: Text answer to the visual question.

Scoring recipe

def compute_accuracy(predictions, gold_answers):
    correct = 0
    for pred, gold in zip(predictions, gold_answers):
        if normalize_answer(pred) == normalize_answer(gold):
            correct += 1
    return (correct / len(gold_answers)) * 100

Common pitfalls

• Few-shot results are averaged over 5 random seeds for demonstration sampling, not a single deterministic run.
• VQA-v2 scores may be inflated for baselines like DFN because MSCOCO (used to train DFN) overlaps with VQA-v2's source data.
• Zero-shot and few-shot evaluations measure different capabilities (direct reasoning vs. in-context learning), so they should not be directly compared without context.

Evidence (verbatim from paper)

Finally, the UniFilter induced MLLM achieves +0.7 and +2.8 average accuracy improvements over the DFN baseline on 4-shot and 8-shot in-context learning, respectively.

Citation

@misc{wang2025unifilter,
  title={Train a Unified Multimodal Data Quality Classifier with Synthetic Data},
  author={Wang et al. (2025)},
  year={2025},
  note={arXiv:2510.15162}
}

• arXiv: 2510.15162