qhjqhj00 mme-realworld-mmbench-eval
Awaker2.5-VL: Stably Scaling MLLMs with Parameter-Efficient Mixture of Experts — Long et al. (2024) (arXiv:2411.10669, 2024)
What this evaluates
Evaluates multimodal large language models on real-world visual perception and reasoning tasks, as well as fine-grained visual-language understanding across multiple dimensions.
Datasets
- MME-Realworld — total 29429; splits: test (-1); repo https://mme-realworld.github.io/home_page.html
- MMBench — total ?; splits: test (-1); repo https://mmbench.opencompass.org.cn/leaderboard
Metrics
Overall score(primary) — range: percent- Percentage of correctly answered questions averaged across all sub-tasks or evaluation dimensions.
Perception score— range: percent- Accuracy on tasks requiring visual recognition and detail extraction.
Reasoning score— range: percent- Accuracy on tasks requiring logical deduction and multi-step visual reasoning.
Input / output format
Input: Image(s) paired with text prompts or questions.
Output: Text responses generated by the model.
Scoring recipe
def compute_metric(predictions, gold_labels):
correct = sum(1 for pred, gold in zip(predictions, gold_labels) if pred == gold)
return (correct / len(gold_labels)) * 100
Common pitfalls
- Competitor results are directly cited from external leaderboards rather than re-run under identical conditions.
- Scores are aggregated averages across diverse sub-tasks, which may mask performance variations on specific capabilities like OCR or spatial reasoning.
Evidence (verbatim from paper)
We conduct evaluation on the latest two multimodal large model benchmarks: (1) MME-RealWorld [[15]]: this benchmark considers images from domains such as autonomous driving, remote sensing, video surveillance, newspapers, street views, and financial charts. It contains 29,429 annotations, covering 43 sub-tasks, with each task having at least 100 questions. (2) MMBench [[12]]: this benchmark is a visual-language model evaluation benchmark developed by the OpenCompass research team. It enables a granular assessment of capabilities ranging from perception to cognition, covering 20 fine-grained evaluation dimensions including object detection, text recognition, action recognition, image understanding, and relational reasoning. All compared models are ranked by the average/overall scores.
Citation
@misc{long2024awaker25vl,
title={Awaker2.5-VL: Stably Scaling MLLMs with Parameter-Efficient Mixture of Experts},
author={Long et al. (2024)},
year={2024},
note={arXiv:2411.10669}
}
- arXiv: 2411.10669