Yihang Chen 陈奕行

We assign each reward greater weight at the denoising timesteps where its feedback is most informative, while preserving user-specified reward budgets.

We introduce the Q-target framework, which unifies supervised fine-tuning methods through target distribution design. Target-SFT balances trust in observed tokens with probability assigned to plausible alternatives.

We propose IRIS, the first framework to improve autoregressive T2I models with reinforcement learning using only an intrinsic reward.

We propose an adaptive diffusion-based purification strategy that can effectively neutralize adversarial corruptions in vision-language models.

We model the multi-step preference alignment problem as a two-player constant-sum Markov game. We propose the MPO, a natural actor-critic algorithm; and the OMPO, an optimistic online gradient descent algorithm.

We introduce the first membership inference attack (MIA) benchmark for vision language models. We also propose the first cross-model MIA pipeline with a new MIA metric.

We study kernel ridge regression in high dimensions under covariate shifts and analyzes the role of importance re-weighting. We also provide asymptotic expansion of kernel functions/vectors under covariate shift.

We propose Order-Preserving GFlowNets (OP-GFNs), which sample composite objects given only the (partial) order. We extend GFlowNets to preference-based optimization in single and multi-objective sampling.

We provide the upper bound of the parameter distribution moving and generalization error on the 0-1 classification task of the infinitely deep and wide ResNets.

We sanity-check prune-at-init methods, and find them hardly exploits any information from the training data. We propose "zero-shot" pruning, which only relies on simple data-independent pruning ratios for each layer.