2 papers across 2 sessions
We propose ADRPO, a method that dynamically adjusts divergence regularization strength based on advantage estimates, enabling more effective fine-tuning of generative models by automatically balancing exploration and exploitation at the sample level.