Research

A review of LOB-Bench, focusing on why generated order book models need rollout-level evaluation beyond one-step loss and stylized facts.

A review of SFT Memorizes, RL Generalizes, focusing on how supervised finetuning and verifier-based reinforcement learning behave under shifted rules and visual inputs.

Modern LLM post-training, from preference learning to reinforcement learning with verifiable rewards.

Training-based routes toward one-step generation for diffusion and flow models, covering distillation, Consistency Models, CTM, MeanFlow, DMD, and Drifting Models.

Study note on path signature in the literature of rough path theory. This post introduces the definition, algebraic structure, and probabilistic interpretation of signatures, bridging rough path theory and modern machine learning, based primarily on *A Primer on the Signature Method in Machine Learning*.

A review of recent research that challenges the necessity of noise level conditioning in generative models, exploring alternative approaches to denoising and flow matching.

A review of researches that accelerate diffusion models in wall clock time by parallelization.

A review of researches that explore test-time scaling for diffusion models.