Real-Time Adaptive Loss Functions for Generative Models Using Reinforcement Learning and Meta-Learning
Aryan Dadwal · 2025
This paper presents a novel approach to training generative models using loss functions that adapt in real time via a meta-learning controller while optionally incorporating a reinforcement-learning (RL) based monitor. Comprehensive ablation studies on a SimpleCNN trained on a 5-image per-class CIFAR-10 subset show that the Meta-Controller-Only configuration yields the best validation accuracy (61.5 %), outperforming a static cross-entropy baseline (60.7 %) by 0.8 percentage points (≈ 1.3 % relative). In contrast, the RL Monitor-Only setting degrades performance (60.4 %), and combining both agents erodes the Meta-Controller's gains due to negative interference (60.6 %). These findings indicate that the learned metarules , rather than the RL policy, are primarily responsible for generalization improvements, and that naïvely fusing the two mechanisms can be counter-productive. The paper details the system architecture, safety mechanisms, experimental protocol, and a critical discussion of the component-wise results and their implications for future adaptive-loss research.