Researchers from UPenn and UIUC introduce Normalizing Trajectory Models (NTM), which use conditional normalizing flows to retain exact likelihood in few-step diffusion sampling. By combining shallow invertible blocks with a deep parallel predictor, NTM achieves state-of-the-art text-to-image results in just four steps while enabling self-distillation through its precise trajectory likelihood.
Termsslug: self-distillation
Self-Distillation
自蒸馏
Process where a model extracts knowledge from its own or a teacher's intermediate representations to optimize itself.
Curated Intelligence
Apple Machine Learning ResearchT1·Models
78 pts
Apple Machine Learning ResearchT1·Models
68 pts
The new RISED framework replaces scalar rewards with LLM-generated structured rubrics to address data selection and lack of intra-group contrast in multi-environment RL. By using textual feedback to guide online data filtering and policy supervision, it achieves the highest mean pass rates across diverse interactive environments.