Mme Realworld Eval

mme-realworld-eval

MM-RLHF: The Next Step Forward in Multimodal LLM Alignment — Zhang et al. (2025) (arXiv:2502.10391, 2025)

What this evaluates

Evaluates multimodal large language models' perception and reasoning capabilities on high-resolution, real-world images across five domains: optical character recognition, remote sensing, diagrams/tables, monitoring, and autonomous driving.

Datasets

• MME-RealWorld — total 29429; splits: test (29429)

Metrics

• Avg (primary) — range: percent
  • Weighted average accuracy across subtasks/domains. Calculated as the sum of per-subtask accuracies multiplied by their respective QA pair counts, divided by the total number of QA pairs.
• Avg-C — range: percent
  • Unweighted average accuracy across subtasks/domains. Calculated as the simple arithmetic mean of per-subtask accuracies.

Input / output format

Input: High-resolution image paired with a text question.

Output: Text response (typically multiple-choice or short answer).

Scoring recipe

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

def compute_avg_accuracy(subtask_accuracies, subtask_sizes):
    weighted_sum = sum(acc * size for acc, size in zip(subtask_accuracies, subtask_sizes))
    return weighted_sum / sum(subtask_sizes)

Common pitfalls

• Closed-source models often compress high-resolution images due to upload size/resolution limits, artificially lowering performance.
• Models may adopt conservative strategies, outputting 'E' (object not present) or refusing to answer due to safety filters, skewing accuracy.
• Weighted vs. unweighted averages can yield significantly different rankings when subtask sizes vary greatly.

Evidence (verbatim from paper)

“Avg” and “Avg-C” indicate the weighted average accuracy and the unweighted average accuracy across subtasks in each domain.

Citation

@misc{zhang2025mmrlhf,
  title={MM-RLHF: The Next Step Forward in Multimodal LLM Alignment},
  author={Zhang et al. (2025)},
  year={2025},
  note={arXiv:2502.10391}
}

• arXiv: 2502.10391