LMMs-Eval: Reality Check on the Evaluation of Large Multimodal Models
(2407.12772)Abstract
The advances of large foundation models necessitate wide-coverage, low-cost, and zero-contamination benchmarks. Despite continuous exploration of language model evaluations, comprehensive studies on the evaluation of Large Multi-modal Models (LMMs) remain limited. In this work, we introduce LMMS-EVAL, a unified and standardized multimodal benchmark framework with over 50 tasks and more than 10 models to promote transparent and reproducible evaluations. Although LMMS-EVAL offers comprehensive coverage, we find it still falls short in achieving low cost and zero contamination. To approach this evaluation trilemma, we further introduce LMMS-EVAL LITE, a pruned evaluation toolkit that emphasizes both coverage and efficiency. Additionally, we present Multimodal LIVEBENCH that utilizes continuously updating news and online forums to assess models' generalization abilities in the wild, featuring a low-cost and zero-contamination evaluation approach. In summary, our work highlights the importance of considering the evaluation trilemma and provides practical solutions to navigate the trade-offs in evaluating large multi-modal models, paving the way for more effective and reliable benchmarking of LMMs. We opensource our codebase and maintain leaderboard of LIVEBENCH at https://github.com/EvolvingLMMs-Lab/lmms-eval and https://huggingface.co/spaces/lmms-lab/LiveBench.
Overview
-
The paper introduces LMMs-Eval, a unified benchmark framework designed to standardize the evaluation of Large Multimodal Models (LMMs) across over 50 tasks.
-
LMMs-Eval Lite addresses evaluation costs by offering a pruned version that balances efficiency and robust evaluation while retaining high correlation with full benchmark results.
-
LiveBench provides a novel approach to evaluating models on real-world data by dynamically updating evaluation datasets from sources like news outlets and forums to ensure relevance and reduced contamination risks.
LMMs-Eval: Reality Check on the Evaluation of Large Multimodal Models
The paper "LMMs-Eval: Reality Check on the Evaluation of Large Multimodal Models" by Zhang et al. introduces a new benchmark framework designed to rigorously evaluate Large Multimodal Models (LMMs). Despite significant advancements in evaluating language models, comprehensive benchmarking of LMMs remains an underexplored area, necessitating a robust and standardized evaluation framework.
Key Contributions
The contributions of this work are threefold:
- LMMs-Eval: This is a unified and standardized benchmark suite that encompasses over 50 tasks across more than 10 models, facilitating comprehensive and apples-to-apples comparison among models.
- LMMs-Eval Lite: To address evaluation cost without sacrificing coverage, this pruned version of LMMs-Eval is designed to balance efficiency and robust evaluation.
- LiveBench: For evaluating models on real-world, dynamically updated data, LiveBench provides a novel low-cost and zero-contamination approach, leveraging continuously updating news and online forum data.
Standardized Framework
LMMs-Eval aims to reduce redundancy and ensure standardized comparisons. Current model evaluations suffer from inconsistent practices (e.g., different data preparation and metrics calculation methods). LMMs-Eval addresses this by providing one command evaluation across multiple models and datasets, standardizing data preprocessing, inference, and metric calculation. As shown in their results, a vast array of models including LLaVA-1.5, Qwen-VL, and InternVL have been evaluated on tasks ranging from captioning in COCO to complex QA in GQA.
Table \ref{tab:fair-comparison} in the paper shows the efficacy of LMMs-Eval in providing standardized comparisons. For instance, LLaVA-NeXT-34B achieves an 80.4% score on RealWorldQA, showcasing competitive performance relative to other models. The framework scales evaluations effectively, reducing manual overhead and ensuring comparability.
Addressing the Evaluation Trilemma
The paper identifies an evaluation trilemma: achieving wide coverage, low cost, and zero contamination concurrently is challenging. While frameworks like Hugging Face’s OpenLLM leaderboard provide broad coverage, they risk contamination and overfitting. Conversely, real-user evaluation platforms like AI2 WildVision are costly due to the need for extensive human interaction data.
LMMs-Eval Lite attempts to navigate these trade-offs by focusing on a smaller yet representative subset of the full benchmark. Using a k-center clustering algorithm to prune data instances, the Lite version retains high correlation with full benchmark results, thereby maintaining reliability while reducing evaluation cost. Figure \ref{fig:cost} illustrates a significant reduction in evaluation time with comparable performance metrics, making it a practical solution for iterative model refinement.
LiveBench: Enhancing Real-World Evaluation
Addressing contamination issues involves not only benchmarking with clean data but also ensuring relevance and timeliness. LiveBench transitions from static to dynamic evaluation sources, gathering data from frequently updated websites like news outlets and forums. This setup tests models on their generalization capabilities more rigorously.
Their analysis reveals high contamination risks in existing benchmarks; for example, datasets like ChartQA show considerable overlap with pretraining data. LiveBench mitigates such risks by dynamically generating up-to-date evaluation datasets and focusing on zero-shot generalization. Monthly curated problem sets ensure continuous relevance.
The evaluation on LiveBench (Table \ref{tab:livebench_results}) reveals that while open-source models have made strides, commercial models such as GPT-4V and Claude-3 still outperform in dynamically updated, real-world scenarios. This underscores the gap between academic benchmarks and practical model utility.
Implications and Future Directions
The implications of LMMs-Eval and its derivatives are significant. For researchers, it provides a robust tool for evaluating model performance across a broad spectrum of tasks under consistent conditions. Practically, it aids in detecting model weaknesses and areas for improvement. Theoretically, the realization of the evaluation trilemma informs ongoing methodological refinements.
Future developments could focus on improving the balance among the trilemma’s three facets or innovating to potentially resolve it. Furthermore, refining contamination detection methods that don't require access to training data could be another avenue of exploration.
In conclusion, LMMs-Eval, along with its Lite and LiveBench variants, offers a comprehensive benchmarking solution for evaluating the growing suite of Large Multimodal Models. As the field continues to evolve, such frameworks will be essential for fostering advancements and ensuring the reliability and robustness of AI systems in diverse real-world applications.
Create an account to read this summary for free: