Assistant Professor, The University of Osaka
2 papers at NeurIPS 2025
We propose Hölder-DPO, the first alignment method with a provable redescending property, which enables robust learning from noisy human feedback by identifying and correcting mislabeled data, improving alignment and model performance.
This study provides an information-theoretic analysis of discrete latent variables in VQ-VAEs, deriving a novel generalization error bound based on the complexity of the latent variables and encoder.
