MetaEvaluator: Cost-Effective Model Evaluation via Meta-Learning
Researchers introduce MetaEvaluator, a model-agnostic framework for evaluating machine learning models on unlabeled data without costly annotation or fine-tuning. It uses meta-learning over a pool of reference models to transfer knowledge, enabling rapid assessment of new models across architectures and modalities. The approach amortizes evaluation costs and eliminates per-model retraining. Experiments demonstrate its effectiveness on unseen datasets. The work is published on arXiv under ID 2605.23595.
Key facts
- MetaEvaluator is a model-agnostic framework for label-free model evaluation.
- It leverages meta-learning over a pool of reference models.
- It enables evaluation of unseen models on unlabeled datasets.
- The framework removes the need for per-model retraining.
- It amortizes evaluation costs across the reference model pool.
- The approach is applicable to diverse architectures and modalities.
- The paper claims to be the first model-agnostic framework for unlabeled evaluation.
- The research is available on arXiv with ID 2605.23595.
Entities
Institutions
- arXiv