4 articles tagged with #sampling-efficiency. AI-curated summaries with sentiment analysis and key takeaways from 50+ sources.
Researchers present LiDAR, a test-time scaling method for diffusion models that improves sample quality alignment with human intent using efficient reward guidance. The approach achieves comparable performance to existing gradient guidance methods while delivering 9.5x faster sampling speeds by computing expected future rewards from marginal samples without neural backpropagation.
Researchers introduce Flux Matching, a generative modeling paradigm that extends beyond score-based models by allowing flexible vector fields with weaker constraints. This advancement enables faster sampling, interpretable models, and dynamics that capture directed variable dependencies while maintaining strong performance on high-dimensional image datasets.
Researchers introduce Pass@K Policy Optimization (PKPO), a reinforcement learning method that optimizes for multiple solution attempts jointly rather than individually, enabling better exploration and problem-solving on harder tasks. The approach derives unbiased estimators for pass@k performance across arbitrary k values and demonstrates improved learning on challenging benchmarks using open-source LLMs.
Researchers introduce Strong Stochastic Flow Maps (SSFMs), a novel framework that extends deterministic flow maps to stochastic differential equations, enabling few-step sampling for diffusion models with pathwise convergence guarantees. The method uses polynomial approximations to Brownian motion and demonstrates improvements over previous approaches in image generation and molecular simulations.
