What type of study is this?

This is a Experimental Study study.

September 24, 2025Open Access

Inference-Time Alignment Control for Diffusion Models with Reinforcement Learning Guidance

View Full Paper

Ask AI

Bookmark

Authors

Inference-time alignment is improved in diffusion models using reinforcement learning, suggesting enhanced control of output distributions.

Reinforcement learning guidance significantly enhances output distribution alignment in diffusion models, improving generative quality.
Dynamic control of the alignment-quality trade-off in RL fine-tuned models is achieved through mathematical equivalence with KL-regularization.
Inference-time method utilizes a geometric average of base and RL fine-tuned outputs, showing superior results across diverse downstream tasks.
Extensive experimentation confirms that this method supports both interpolation and extrapolation, offering practical flexibility for generative alignment.

Jin et al. (2025) studied this question.

View Full Paper

Ask AI

Bookmark

Inference-time alignment is improved in diffusion models using reinforcement learning, suggesting enhanced control of output distributions.