Mudaif Vl Eval

mudaif-vl-eval

Optimizing Vision-Language Interactions Through Decoder-Only Models — Tanaka et al. (2024) (arXiv:2412.10758, 2024)

What this evaluates

Evaluates a decoder-only vision-language model's ability to perform visual question answering, image captioning, and multimodal reasoning. It measures cross-modal alignment, computational efficiency, and robustness to input variations like resolution and noise.

Datasets

• VQA-v2 — total ?; splits: test (-1)
• GQA — total ?; splits: test (-1)
• VizWiz — total ?; splits: test (-1)
• SEED — total ?; splits: test (-1)
• MM-Vet — total ?; splits: test (-1)

Metrics

• accuracy (primary) — range: percent
  • Percentage of correctly predicted answers out of the total number of questions.
• BLEU — range: [0, 1]
  • Standard n-gram overlap metric for evaluating image captioning quality.
• SEED Score — range: other
  • Composite score for multimodal reasoning benchmarks.
• MM-Vet Score — range: other
  • Composite score evaluating multimodal capabilities and instruction following.

Input / output format

Input: Paired image and text input (question or instruction prompt).

Output: Text string containing the predicted answer or caption.

Scoring recipe

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

Common pitfalls

• SEED and MM-Vet use proprietary or complex composite scoring rubrics that are not publicly detailed in this paper; users must use official evaluation scripts.
• BLEU is reported for captioning but the specific dataset and n-gram order are not specified, requiring standard implementation assumptions.
• Human evaluation uses a 1-5 Likert scale averaged across tasks, which may introduce rater bias and lacks inter-annotator agreement metrics.

Evidence (verbatim from paper)

Metrics include accuracy for VQA, BLEU for captioning, and composite scores for multimodal benchmarks. Human evaluation assesses relevance, coherence, and informativeness.

Citation

@misc{tanaka2024mudaif,
  title={Optimizing Vision-Language Interactions Through Decoder-Only Models},
  author={Tanaka et al. (2024)},
  year={2024},
  note={arXiv:2412.10758}
}

• arXiv: 2412.10758