5 appearances on the backlist front page in the last 30 days.
Two insights from LeapAlign: 1. Gradient descent, rather than GRPO, is native to diffusion post-training. 2. Early generation steps should be trained, such that image layout can be better optimized. Thanks @hillbig for posting this work.
Across 21 recent diffusion methods, improvements on ImageNet did not predict text-to-image improvements under identical DiffusionBench settings
Diffusion is differentiable. LLMs aren't. So why is the diffusion community copying RL methods (GRPO etc.) from LLMs? The native post-training for diffusion is gradient descent such as ReFL and LeapAlign. Paper: http:// arxiv.org/abs/260
Diffusion is differentiable. LLMs aren't. So why is the diffusion community copying RL methods (GRPO etc.) from LLMs? The native post-training for diffusion is gradient descent such as ReFL and LeapAlign. Paper: http:// arxiv.org/abs/260
update: tested GATS on GPT-5.5 BFCL: +5.34% τ²-bench: +2.32% so there is consistent gain of three GPT models: GPT-4o, GPT-5, and GP5.5. simulated feedback for tool-calls refinement keeps working even as base models get stronger. code