Adaptive Meta-Learning with Reinforcement-Guided Task Weighting in an Enhanced Reptile Framework

Abstract

Meta-learning algorithms aim to endow models with the ability to learn new tasks rapidly from few examples. Reptile, a first-order meta-learning method, achieves this by searching for an initialization that is close to many tasks. Recent work improved Reptile through multi stage gradient aggregation and adaptive inner-loop learning rates, yet task weighting remained static. We propose Reinforced Reptile R2, an enhanced Reptile variant in which a lightweight reinforcement learning (RL) agent assigns dynamic weights to tasks during meta-training. The agent observes task specific performance signals and outputs continuous weights that modulate the meta-gradient. Extensive experiments on four vision benchmarks (Omniglot, mini-ImageNet, tiered-ImageNet, CIFAR-FS) and one NLP benchmark (SNLI-FewShot) show that R2 outperforms classical Reptile by up to 4.1% in 5-way 1-shot accuracy while reducing convergence time by 18%. Ablation studies confirm that RL guided task weighting accelerates learning on difficult tasks without overfitting to easy ones. R2 offers a simple, resource-efficient drop-in replacement for existing Reptile pipelines.