Tag
The paper identifies that distilled few-step diffusion models suffer diversity collapse due to activation variance concentrating into fewer directions in an intermediate block, and proposes ReDiverse — an adapter fine-tuning method (ReDiverse-FT) and a training-free activation steering mechanism (ReDiverse-TF) — to restore sample diversity while preserving generation quality across image and video distillation benchmarks.
This paper introduces FADE (Focal Advantage with Dynamic Entropy), a self-adapting advantage function that dynamically schedules gradient weights during RL post-training of LLMs, achieving faster convergence and better accuracy-diversity trade-offs compared to static baselines.