Xujiang Zhao

I am a researcher at NEC Laboratories America. I received the Ph.D. in Computer Science Department at The University of Texas at Dallas, in 2022. My advisor is Prof. Feng Chen. I also got an MS at the University of Science and Technology of China (USTC) and a bachelor at Chongqing University in China.

Email  /  CV  /  Google Scholar  /  LinkedIn  /  Github

" The fear of the LORD is the beginning of wisdom " - Psalms 111:10

I'm interested in machine learning, NLP, and data mining, especially in Uncertainty Quantification and Reasoning, Large Language Models, Reinforcement Learning, Natural Language Processing, and Graph Neural Networks.

We theoretically analyze the failure case and reason for treating LLM detection tasks as binary classification tasks and propose to transform the task to an out-of-distribution detection task.

We introduce SolverLLM, a training-free framework that leverages test-time scaling to solve diverse optimization problems.

In this paper, we proposed a novel framework that propagates uncertainty through each step of an LLM-based agent’s reasoning process.

We present a comprehensive survey on domain specification techniques for large language models, an emerging direction critical for large language model applications. This paper was cited by 2024 Economic Report of the President of the United States

In this paper, we have shown that framing code generation as an optimization task over the code space and applying SCATTERED FOREST SEARCH is highly effective..

In this paper, we proposed MixLLM, a dynamic routing system that selects the most suitable LLM for each query by balancing response quality, cost, and latency.

In this paper, we first derive a unified view of prompt tuning and then present a novel dynamic prompting approach that can significantly improve the performance of prompt tuning while adding only a few additional parameters..

In this paper, we introduce a novel evidential method, Multi-Label Evidential Graph Neural Networks, to predict uncertainty for multiple classes on graph data.

In this paper, we present a comprehensive exploration of strategies for fine-tuning Large Language Models (LLMs) to incorporate domain-specific knowledge while upholding data privacy.

This paper has proposed a novel framework, DCFDG, to address issues of fairness within continuously evolving dynamic environments.

We introduce D-PRUNER, an innovative unstructured dual-pruning method for domain-specific compression on LLM. It is able to extract a compressed, domain-specific, and task-agnostic LLM by identifying weights that are pivotal for both generality and specificity.

We provide an Uncertainty Quantification and Decomposition of In-Context Learning of Large Language Model.

We present an off-the-shelf framework KEEP to predict answers for open-ended commonsense reasoning without requiring answer candidates and a pre-defined answer scope.

This paper aims to explore spurious relationships in structural causal models (SCMs) that arise due to sensitive factors.

We propose a Graph Edge Re-weighting via Deep Q-learning (GERDQ) framework to calibrate the graph neural networks.

We study the problem of early event detection in multi-lable classification, and propose a novel framework, Multi-Label Temporal Evidential Neural Network (MTENN), for multi-label uncertainty estimation in temporal data.

In this survey paper, we provide a comprehensive technical review of the existing knowledge-enhanced reasoning techniques across the diverse range of application domains.

In this survey paper, we study the mature uncertainty research in belief/evidence theories in machine learning/deep learning to tackle complex problems under different types of uncertainty.

We study the key causes about the negative impact of OODs (boundary OODs and faraway ODDs) on SSL and proposed a simple unified robust SSL approach for many existing SSL algorithms in order to improve their robustness against OODs.

In this paer, we investigate automated GNN calibration by marrying uncertainty estimation to the hyperparameter optimization (HPO) problem.

We propose a novel Polyphonic Evidential Neural Network to model the evidential uncertainty of the class probability with Beta distribution to solve the sound event early detection problem.

In this paper, we proposed a novel layer-adaptive OOD detection framework (LA-OOD) for DNNs that can fully utilize the intermediate layers’ outputs.

We propose a subset selection algorithm in semi-supervised learning to speed up the SSL training. In addition, this algorithm achieve a better performance when unlabeled data consists of Out-of-Distribution (OOD) data and imbalance.

In this work, we propose a self-learning framework that further utilizes unlabeled data of target languages, combined with uncertainty estimation in the process to select high-quality silver labels..

CLEAR: Contrastive-Prototype Learning with Drift Estimation for Resource Constrained Stream Mining.

By considering the multidimensional uncertainty, we proposed a novel uncertainty-aware evidential NN called WGAN-ENN (WENN) for solving an out-of-distribution (OOD) detection problem.

A multi-source uncertainty framework of GNN that reflecting various types of uncertainties in both deep learning and belief/evidence theory domains for node classification predictions.

A novel multi-source uncertainty prediction approach that enables deep learning models to be actively trained with much less labeled data.

we propose a highly scalable opinion inference probabilistic model, namely Adversarial Collective Opinion Inference (Adv-COI), which provides a solution to infer unknown opinions with high scalability and robustness under the presence of uncertain,.

This paper presents a new approach, called a regularized EEvidential Neural Networks (ENN), that learns an ENN based on regularizations related to different characteristics of inherent data uncertainty.

This paper proposed three DRL-based schemes combining Subjective Logic and deep reinforcement learning where a reward is given based on a different type of uncertainty (i.e., vacuity, dissonance, or monosonance).

GCN-GRU:Combine Graph Convolutional Network (GCN) and the Gated Recurrent Units (GRU) to model the topological and temporal heterogeneous dependency information of a given dynamic network and conflicting opinions based on robust statistics (uncertainty)

GCN-VAE-SL: The proposed DL-based opinion inference model handles node-level opinions explicitly in a large-scale network using graph convoluational network (GCN) and variational autoencoder (VAE) techniques.

We propose a general framework to model and infer heterogeneous uncertainty information in network data based on GCN and node-level opinions via knowledge distillation.

Design and source code from Jon Barron's website