Tianyi Zhang /<br>张天怡

Tianyi Zhang /
张天怡

Expertise in Natural Language Processing,
Cognitive Science

Background

Pursuing one’s dreams is rarely smooth, but I am deeply grateful to those who have supported me along the way. With your support and my continued effort, I believe my dream can become a reality.

I currently have no connection with a professor at UPenn. Please do not share or communicate any information about me to them without my explicit consent. This behavior has had a serious and ongoing adverse impact on my professional opportunities. I have shared my experience in the document. If this behavior continues, I may identify the individual involved.
This statement is intended to protect myself and is not directed against any individual. I appreciate your understanding and respect for my request.

I have extensive experience in language model reasoning. My work includes knowledge graph–based retrieval-augmented generation systems, language model continual pretraining, neural-symbolic reasoning, event schema induction, and synthetic data augmentation. I have collaborated with many excellent professors in the U.S. and Europe, and I am very grateful for their recognition and support.
To know more about my research experience, please check my CV, SOP and recommendation letters.

My research connects human cognition theories with NLP through:

  1. Cognitive-Inspired Loop Pretraining:
    In Continual Pretraining[1], I demonstrate that modeling the human NL → KG → NL learning process as a looped pretraining task facilitates performance on downstream knowledge-intensive tasks such as summary, QA, and NLI.
  2. Zone of Proximal Development (ZPD):
    In Proc2PDDL[2], I show that instructing LLMs to incrementally build required skills — aligned with Vygotsky’s ZPD — can effectively support complex Text-to-Code translation tasks.

My research interests include:

  1. Reasoning in Natural/Symbolic Language
  2. Language Model Pretraining
  3. Interdisciplinary in NLP, CV, and Robotics

Publications

  1. Effective Domain Adaptation of Instruction-Tuned LLMs for Knowledge-Intensive Tasks. In submission 2025
    Zhang, T.*, Mai, F.*, Flek, L.
    paper
  2. PROC2PDDL: Open-Domain Planning Representations from Texts. NLRSE@ACL 2024
    Zhang, T.*, Zhang, L.*, Hou, Z., Wang, Z., Gu, Y., Clark, P., Callison-Burch, C., and Tandon, N.
    paper poster oral
  3. PDDLEGO: Iterative Planning in Textual Environments. *SEM 2024
    Zhang, L., Jansen, P., Zhang, T., Clark, P., Callison-Burch, C., Tandon, N.
    paper oral
  4. WorldWeaver: Procedural World Generation for Text Adventure Games. Wordplay@ACL 2024
    Jin, M., Kaul, M., Ramakrishnan, S., Jain, H., Chandrawat, S., Agarwal, I., Zhang, T., Zhu, A., Callison-Burch, C.
    paper
  5. Human-in-the-Loop Schema Induction. ACL Demo 2023
    Zhang, T.*, Tham, I. *, Hou, Z. *, Ren, J., Zhou, L., Xu, H., Zhang, L., Martin, L., Dror, R., Li, S., Ji, H., Palmer, M., Brown, S., Suchocki, R., and Callison-Burch, C.
    paper poster oral
  6. Question-Answering Data Augmentation for Argument Role Labeling. 2022
    paper

Education