Researchedit

Xinbao Qiao's current research focuses on data-centric ML and AI and networks, including AI for Networks and Networks for AI.

Research thesisedit

Qiao's work primarily studies lifecycle management of data in AI models, focusing on theoretical methods and practical problems that arise as data are acquired, used, and deleted. The related work aims to improve the reliability, interpretability, and controllability of AI models in heterogeneous, computation-constrained, and communication-constrained environments.

1. In data acquisition, it studies synthetic data generation and its effects on quality, privacy, and generalization.
2. In data use, it focuses on data modeling, collaborative optimization, and system design in distributed/federated learning, AI for Networks, and Networks for AI.
3. In data deletion, it studies machine unlearning and data influence evaluation, exploring how to preserve model performance while protecting privacy and satisfying deletion requests.

AI and networksedit

AI and Networks covers the intersection of AI with networking and communication systems: AI for Networks, Networks for AI, decentralized learning, data pruning, and collaborative evaluation. In the current CUHK doctoral stage, this line is paired with data-centric ML and includes distributed tools such as Wasserstein barycenters, where multiple local distributions can be combined into a shared distributional reference without treating raw-data pooling as the default assumption.

The under-review manuscript Decentralized Free-Support Wasserstein Barycenter develops this direction through free-support barycenter computation: learning support locations lets the shared reference adapt to distributional geometry, while separating the communication needed for barycenter quality from that needed for network agreement.

Machine unlearningedit

Machine Unlearning studies certified data removal and low-cost update mechanisms after deletion requests. Related work includes Hessian-Free Online Certified Unlearning, Beyond Binary Erasure: Soft-Weighted Unlearning for Fairness and Robustness, DynFrs: An Efficient Framework for Machine Unlearning in Random Forest, Influence Functions, and Certified Data Removal.

Synthetic dataedit

Synthetic Data studies recursive synthetic-data training, Data Selection, Sample Selection Bias, Model Collapse, and collaborative mitigation in low-resource data silos. The central paper is When Sample Selection Bias Precipitates Model Collapse, which frames model collapse as especially risky when real-data coverage is scarce or fragmented.

LLM reliabilityedit

LLM Reliability studies whether model responses remain grounded in adequate evidence. The NeurIPS 2026 paper Illusory Pattern Perception Drives Spurious Inference in Large Language Models uses three psychological tasks to compare illusory pattern perception in LLMs and humans.

Data centric ML and trustworthy AIedit

Data Centric ML covers data selection, valuation, filtering, and evaluation. Trustworthy AI connects unlearning, fairness, robustness, privacy, security, interpretability, and reliability.

Geometry and distributed learningedit

Wasserstein Geometry, Distributed Wasserstein Barycenter, and Distributed Learning provide tools for collaborative evaluation, optimal-transport proxies, decentralized data access, and distributional references for networked AI systems.