Yijun Liao
Yijun Liao

Yijun Liao

First-year Undergraduate · Chongqing University of Technology

My research interests lie in understanding how language models learn, adapt, and make decisions. I work across LLM agents, reinforcement learning and alignment, and mechanistic interpretability, with an emphasis on controlled experiments, causal analysis, and rigorous evaluation. I am especially interested in the gap between what a model achieves and the mechanisms or strategies that produce it: strong task performance alone neither establishes robust generalization nor explains how a model reaches its decisions.

liuyingliao0620@gmail.com GitHub Google Scholar

Research Interests

RL & Alignment

  • Preference optimization and LLM post-training
  • Reinforcement learning for language models
  • Alignment in dynamic, long-horizon contexts

LLM Agents

  • Credit assignment and failure attribution
  • Learning from interaction and experience
  • Evaluation and generalization of agent behavior

Mechanistic Interpretability

  • Causal interventions and circuit analysis
  • Training dynamics and mechanism emergence
  • In-context learning and knowledge conflicts

News

Publications & Preprints

Fixed-offset positional shortcut versus retrieval of the current value Preprint

Shortcut Before Circuit: Document Statistics Time In-Context Conflict Resolution

Yijun Liao, Fanwei Liang

Under review, 2026

We train 26M-parameter transformers on a synthetic language where recency and rarity always co-occur, then pull the two cues apart with a minimal edit. Models reach near-perfect accuracy, yet their behavior under intervention is inconsistent. In our controlled setting, document statistics influence when models move beyond positional shortcuts, while the mechanisms that emerge vary across random seeds.

arXiv PDF 
@article{liao2026shortcut,
  title   = {Shortcut Before Circuit: Document Statistics Time In-Context Conflict Resolution},
  author  = {Liao, Yijun and Liang, Fanwei},
  journal = {arXiv preprint arXiv:2608.24460},
  year    = {2026}
}
Dual-zone temporal attention: anchor zone kept, older history down-weighted Preprint

Overcoming State Inertia: Minimally Invasive Temporal Alignment for Evolving Contexts

Yijun Liao

arXiv preprint, 2025

Long-context dialogue models cling to stale history and fail to follow shifting user intent, while existing fixes erode general language ability. We propose DZ-TiDPO, combining conflict-aware preference optimization with a temporal attention bias that decouples state updates from language modeling, and introduce IC-Bench for evaluation.

arXiv PDF Code 
@article{liao2025overcoming,
  title   = {Overcoming State Inertia: Minimally Invasive Temporal Alignment for Evolving Contexts},
  author  = {Liao, Yijun},
  journal = {arXiv preprint arXiv:2512.03704},
  year    = {2025}
}

Academic Service

  • 2026Program Committee Member, AAAI 2027 (AAAI Conference on Artificial Intelligence)

Education

  • 2026 – nowChongqing University of Technology, Undergraduate

© Yijun Liao · Last updated Oct 2026