Xueqiang (Patrick) Xu

I am a first-year CS PhD student at UIUC, where I am advised by Prof. Jiawei Han. I completed my undergraduate studies at UIUC, graduating with a Highest Honors B.S. in Computer Science. I work with Shi Feng this summer on LLM safety alignment.

徐学强  /  Email  /  Google Scholar  /  GitHub  /  LinkedIn  / 

[Jun 2026] 🚀 New preprint: Harness-1, a 20B search agent trained with RL inside a state-externalizing harness — it matches frontier-model searchers on agentic search while staying fully open-source (code).
[May 2026] 🎉 Two papers accepted: one on LLM Reranking at ICML 2026, and one on Compact LLM Reranking at KDD 2026.
[Jan 2026] One paper on zero-shot entity structure extraction ZOES has been accepted by EACL 2026 Main Conference.
[Dec 2025] We released our paper on Adaptation of Agentic AI, with public repository here. Hope you enjoy reading it!.
[Aug 2025] 🔥 Two papers accepted to EMNLP 2025: one in s3: Training Search Agent via RL, and one in LogiCoL: Logically-Informed Contrastive Learning for Set-based Dense Retrieval.

My research focuses on building trustworthy and aligned LLM agent systems. I study how to design, train, and evaluate language model agents so that they can continually learn from experience, reason reliably, and remain aligned with human intent in complex real-world tasks. I approach this goal through two interconnected directions:

• Agent Continual Learning — enabling LLM agents to perform robust continual learning from memory and external knowledge. I investigate how agents can accumulate, organize, and reuse experience through mechanisms such as structured memory, retrieval, and reinforcement learning, so that they improve over time rather than treating each task in isolation.
• LLM Safety and Alignment — studying how to make LLMs honest, harmless, and steerable through alignment techniques such as outcome-reward RL, control vectors, and interpretability-driven evaluation. I explore mechanisms to detect, prevent, and mitigate unsafe or unintended behaviors as models become increasingly capable and agentic.

Preprint

We introduce ZOES, a novel approach to entity structure extraction that does not require any schema or annotated samples. ZOES operates via a principled mechanism of enrichment, refinement, and unification, based on the insight that an entity and its associated structure are mutually reinforcing.

Preprint Code

In this work, we propose s3, a lightweight, model-agnostic framework that decouples the searcher from the generator using only 2.4k data in the RL training process.

arXiv Code

We unify the rapidly expanding research landscape of agentic AI systems into a systematic framework that spans both agent adaptations and tool adaptations. Cutting-edge agentic AI systems are built on foundation models that can be adapted to plan, reason, and interact with external tools to perform increasingly complex and specialized tasks.

Preprint Code

We propose TELEClass, which combines the general knowledge of LLMs and task-specific features mined from an unlabeled corpus. TELEClass automatically enriches the label taxonomy with class-indicative features and utilizes novel LLM-based data annotation and generation methods specifically tailored for hierarchical text classification.

Preprint

We introduce LogiCoL, a logically-informed contrastive learning objective for dense retrievers that handles queries with logical connectives. LogiCoL learns to respect subset and mutually-exclusive set relations between query results via soft constraints expressed through t-norm, achieving improvements in both retrieval performance and logical consistency.

• City Scholar at UIUC
• Illinois Scholars Undergraduate Research
• IIDAI scholar

Template from Jon Barron