Vqa Benchmarks Eval

vqa-benchmarks-eval

2.5 Years in Class: A Multimodal Textbook for Vision-Language Pretraining — Wenqi Zhang et al. (2025) (arXiv:2501.00958, 2025)

What this evaluates

Evaluates vision-language models' ability to answer questions across general knowledge, OCR, mathematics, and science domains using few-shot in-context learning. It also probes the model's capacity to attend to interleaved image-text contexts through a 'cheat test' protocol.

Datasets

• TextVQA — total ?; splits: test (-1)
• OKVQA — total ?; splits: test (-1)
• MathVista — total ?; splits: test (-1)
• MathVision — total ?; splits: test (-1)
• MathVerse — total ?; splits: test (-1)
• ScienceQA-IMG — total ?; splits: test (-1)

Metrics

• accuracy (primary) — range: percent
  • Percentage of correctly predicted answers out of total test instances.
• perplexity — range: other
  • Exponential of the average negative log-likelihood of the ground truth tokens.

Input / output format

Input: Interleaved sequence of images and text (questions, and optionally few-shot examples with images, questions, and answers).

Output: Textual answer string.

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 examples are randomly sampled or retrieved, which can introduce variance in accuracy scores across runs.
• The 'cheat test' protocol replaces a few-shot example with the actual test sample, measuring in-context attention rather than generalization, which may inflate scores on benchmarks requiring complex reasoning.

Evidence (verbatim from paper)

Following OpenFlamingo*[[5]]* and OmniCorpus*[[25]], we evaluate the performance of the pre-trained models on two VQA benchmarks (TextVQA[[43]], OKVQA[[35]]), three visual reasoning benchmarks (MathVista, MathVision, MathVision), and ScienceQA-IMG[[34]], covering general, OCR, mathematics, and science domains. We compute model accuracy in few-shot settings using either randomly sampled or retrieved examples as previous works[[21], [52], [25]]*.

Citation

@misc{zhang2025multimodaltextbook,
  title={2.5 Years in Class: A Multimodal Textbook for Vision-Language Pretraining},
  author={Wenqi Zhang et al. (2025)},
  year={2025},
  note={arXiv:2501.00958}
}

• arXiv: 2501.00958