Textual Equilibrium Propagation for Deep Compound AI Systems
ICLR 2026
Summary A local learning principle for optimizing deep compound AI systems that mitigates exploding and vanishing textual gradients in long-horizon workflows.
Can Textual Gradient Work in Federated Learning?
ICLR 2025
Summary We systematically explore the potential and challenges of incorporating textual gradient into Federated Learning, introducing FedTextGrad - a novel FL paradigm for optimizing LLMs.
Revisiting Delta-Parameter Pruning For Fine-Tuned Models
ICLR 2025
Summary A comprehensive study of delta-parameter pruning that introduces DARq and AdamR to address the limitations of existing methods, enabling efficient storage and deployment of multiple fine-tuned models.
Local Superior Soups: A Catalyst for Model Merging in Cross-Silo Federated Learning
NeurIPS 2024
Summary An innovative model interpolation-based local training technique that enhances local training across different clients through regularized model interpolation, acting as a catalyst for seamless adaptation of pre-trained models in federated learning.
Forgettable Federated Linear Learning with Certified Data Removal
WWW 2024 FL@FM Workshop (Best Paper Award) 2024
Summary A federated learning framework that enables certified data removal through linear approximation and efficient removal strategies, providing theoretical guarantees for the right to be forgotten.
Debiased Noise Editing on Foundation Models for Fair Medical Image Classification
MICCAI 2024
Summary A novel debiasing strategy that mitigates spurious correlations in medical images through noise editing, enabling fairness-aware image processing for both white-box and black-box foundation model APIs.
FedSoup: Improving Generalization and Personalization in Federated Learning via Selective Model Interpolation
MICCAI 2023
Summary A novel federated model soup method that optimizes the trade-off between local and global performance through selective interpolation of model parameters, alleviating overfitting and seeking flat minima for improved generalization.
VITA: A Multi-Source Vicinal Transfer Augmentation Method for Out-of-Distribution Generalization
AAAI 2022
Summary A novel data augmentation method that generates diverse on-manifold samples through multi-source vicinal transfer to improve model robustness against image corruptions.
Benchmarks for Corruption Invariant Person Re-Identification
NeurIPS 2021
Summary The first comprehensive study establishing five ReID benchmarks for learning corruption invariant representations, providing insights on robustness of transformer vs CNN models and cross-dataset generalization.