Pengfei Yu

life4pf AT gmail DOT com

Working on Agentic RL for LLMs.

PhD in Computer Science

University of Illinois Urbana–Champaign University of Illinois Urbana–Champaign.

BEng in Electronic Engineering; BSc in Mathematics

Tsinghua University Tsinghua University

Publications here are selected and ranked based on my personal preference instead of citations or venues. For a complete list, please visit my Google Scholar Profile.

Information Association for Language Model Updating by Mitigating LM-Logical Discrepancy
–Pengfei Yu, Heng Ji (CoNLL 2024) paper

I completed the first version of this project in 2023, right after ChatGPT was released. From today’s perspective, the experiments are not as comprehensive as I would like. Although the paper didn’t land at a top venue, I still like its central arguments:

  • Learning new knowledge involves (1) being able to recall it when needed and (2) applying it appropriately. This work suggests that (1) is often the harder part.
  • LLMs are not logical computers: optimizing $P(y \mid x) \to 1$ does not—and arguably should not—establish a logical knowledge $x \to y$ in the model.

EVEDIT: Event-based Knowledge Editing for Deterministic Knowledge Propagation
–Jiateng Liu, Pengfei Yu, Yuji Zhang, Sha Li, Zixuan Zhang, Ruhi Sarikaya, Kevin Small, Heng Ji (EMNLP 2024) paper

This work continues the line of thinking above, but with a novel finding on LLM training, which, in some ways, behaves surprisingly “logically.” When training data contains logical inconsistencies about a topic, the model becomes uncertain and inconsistent in its responses, just like a confused person who has been given contradictory information.

The Law of Knowledge Overshadowing: Towards Understanding, Predicting and Preventing LLM Hallucination
–Yuji Zhang, Sha Li, Cheng Qian, Jiateng Liu, Pengfei Yu, Chi Han, Yi R. Fung, Kathleen McKeown, ChengXiang Zhai, Manling Li, Heng Ji (ACL 2025 Findings) paper

Most of the credit for this work goes to Dr. Yuji Zhang. The part I found especially compelling (and contributed to) is the motivating idea that language models can form spurious associations between co-occurring concepts: when “A X B” appears frequently, A can become a strong cue for producing B later, regardless of X. The paper also supports an intuition I’ve had for a while: hallucination is a form of generalization.