Search papers, labs, and topics across Lattice.
Tencent Hunyuan
5
0
6
Early tokens in LLMs can lead to compounding errors, but CPPO's position-sensitive approach offers a solution that boosts reasoning accuracy and training stability.
Flow-DPPO outperforms traditional PPO methods by achieving higher rewards and greater training stability through a novel divergence proximal constraint.
FlowBP redefines reward backpropagation by transforming the backward trajectory into a design object, leading to significant improvements in model alignment with human preferences.
Smooth gradient adjustments in DRPO prevent harmful policy shifts, leading to more stable and efficient LLM training.
Forget slow, complex training: you can now distill diffusion models to just 4 steps and still beat the state-of-the-art in preference alignment, aesthetics, and composition.
