Language models are increasingly used in tasks ranging from software development to scientific research. These tasks often require models to explain how evidence supports a conclusion, decide which information matters, and build on earlier findings. My research focuses on the structure of explanations and reasoning. I use this structure to evaluate and improve how AI systems reason and learn.

• Understanding and preserving context. I develop frameworks based on Questions Under Discussion (QUD) that model relationships across a text through questions and answers. For agents, I build memory that preserves earlier findings along with the evidence and reasoning behind them. Selected work: QUD parsing (Findings of ACL 2023); ContextWeaver (under submission, 2026); ContextGraph (in progress)
• Evaluating understanding and generation. I evaluate whether a model's interpretation is supported by its context. This evaluation reveals interpretation errors that common metrics miss. I also measure structural repetition in model outputs. Selected work: QUDeval (EMNLP 2023); QUDsim (COLM 2025)
• Learning what information matters. I use questions to separate what to explain from how to explain it. I predict which questions are important from their context, and I extend this framework to scientific figures. Selected work: QSalience (EMNLP 2024, Outstanding Paper Award); ElabQUD (EMNLP 2023); Multimodal QUD (under submission, 2026)

Ongoing and future research

• Learning efficiently from experience. QUDsim showed that models reuse the same discourse structures across topics. I study when such repetition helps learning and when it wastes computation, in training data and across agents.
• Collaboration between people and AI. Building on my evaluation work, I study how AI systems interpret and respond to human feedback. I also study how agents follow limits on their actions.
• AI for scientific discovery. I extend my work on question importance to scientific investigations, where the value of a question changes as new evidence arrives. I study how AI systems choose observations and revise research plans in light of the results. Ongoing work with CosmicAI in astronomy provides an initial setting.

Education

Ph.D., (Aug. 2022 - Now) in Computer Engineering

M.S., (Jan. 2020 - Aug. 2022) in Computer Engineering

Department of Electrical and Computer Engineering,

University of Texas at Austin.
Advisors: Prof. Jessy Li, Prof. Alex Dimakis
Group: UT NLP Group, Wireless Networking and Communications Group(WNCG)

B.Eng., (Spt. 2014 - Jul. 2019) in Computer Science and Technology

B.A., (Spt. 2014 - Jul. 2019) in Japanese

Department of Computer Science
Dalian University of Technology.

Undergraduate Exchange Student, (Spt. 2017 - Aug. 2018) in Computer Science

Department of Information & Communication Engineering,
The University of Tokyo.
Advisor: Prof. Toshihiko Yamasaki
Group: Aizawa-Yamasaki Laboratory (now Aizawa-Yamakata-Matsui Laboratory)

Selected Publications

^†Indicates equal contribution.

• Multimodal QUD: Inquisitive Questions from Scientific Figures [project website] [dataset]
  Yating Wu, William Rudman, Venkata S Govindarajan, Alexandros G. Dimakis, Junyi Jessy Li
  Under submission, 2026.

• ContextWeaver: Selective and Dependency-Structured Memory Construction for LLM Agents [project]
  Yating Wu, Yuhao Zhang, Sayan Ghosh, Sourya Basu, Anoop Deoras, Jun Huan, Gaurav Gupta
  Under submission, 2026.

• QUDsim: Quantifying Discourse Similarities in LLM-Generated Text [project] [code and data] [paper]
  Ramya Namuduri, Yating Wu, Anshun Asher Zheng, Manya Wadhwa, Greg Durrett, Junyi Jessy Li
  Conference on Language Modeling (COLM), 2025.

• Which questions should I answer? Salience Prediction of Inquisitive Questions [project] [code and data] [paper]
  Yating Wu^†, Ritika Mangla^†, Alexandros G. Dimakis, Greg Durrett, Junyi Jessy Li
  Conference on Empirical Methods in Natural Language Processing (EMNLP), 2024. (Oral)
  Outstanding Paper Award 🏅

• QUDeval: The Evaluation of Questions Under Discussion Discourse Parsing [project] [code and data] [paper]
  Yating Wu, Ritika Mangla, Greg Durrett, Junyi Jessy Li
  Conference on Empirical Methods in Natural Language Processing (EMNLP), 2023. (Oral)

• Elaborative Simplification as Implicit Questions Under Discussion [project] [code and data] [paper]
  Yating Wu^†, William Sheffield^†, Kyle Mahowald, Junyi Jessy Li
  Conference on Empirical Methods in Natural Language Processing (EMNLP), 2023.

• Discourse Analysis via Questions and Answers: Parsing Dependency Structures of Questions Under Discussion [project] [code] [paper]
  Wei-Jen Ko, Yating Wu, Cutter Dalton, Dananjay Srinivas, Greg Durrett, Junyi Jessy Li
  Findings of the Association for Computational Linguistics (ACL), 2023.

• Modeling Bilingual Disfluencies with Large Language Models [project]
  Negin Raoof^†, Yating Wu^†, Carlos Bonilla^†, Junyi Jessy Li, Stephanie M Grasso, Alexandros G. Dimakis, Zoi Gkalitsiou
  ICML Workshop on LLMs and Cognition, 2024.

Awards

• MIT EECS Rising Stars 2025
• Outstanding Paper Award in EMNLP 2024
• VMware Codehouse Palo Alto 1st place, Jul. 2021
• Outstanding graduates, Jun. 2019
• Outstanding graduates, Jun. 2014

Experience

Working Experience

• AWS AI Labs, Summer 2025
  Applied Scientist Intern. I worked on long-context memory support for LLM coding agents.
• Amazon Alexa AGI, Summer 2024
  Applied Scientist Intern. I worked on research problems focused on enhancing the planning and tool-using capabilities of LLM agents.
• Amazon Prime Video, Summer 2023
  Software Develop Engineer Intern. I implemented an in-game notification system.
• Amazon Prime Video, Summer 2021
  Software Develop Engineer Intern. I implemented a Java-based ranking system for live events. The project got launched in prime video offical site.

Teaching Experience

• CS391L: Machine Learning (Head TA). Summer 2022 - Fall 2025.
• CS391L: Machine Learning (TA). Fall 2021 - Summer 2022
• EE422C: Software Design & Implementation II (TA). Summer 2020 - Spring 2021

Mentoring

• Ritika Mangla, Master's student, UT Austin, 2022-2024. Co-authored two EMNLP papers, including an Outstanding Paper Award.
• Ashwin Ram, Master's student, UT Austin, 2024.
• Tianyang Xu, Ph.D. student, TTIC.

Services

• Workflow Chair, EMNLP 2026
• Reviewer: ACL ARR 2024-2025, ACL 2024-2025, EMNLP 2024, COLM 2024-2025, ICML 2025 XAI4Science, AAAI 2025 MARW, NeurIPS 2024 Behavioral ML, ICML 2024 DMLR, ICML 2024 LLMs and Cognition

RESEARCHResearch DeltaUnroll ↔Roll up ↔

Questions under discussion

2023QUD parsingDiscourse structure 2023QUDevalEvaluating the questions 2025QUDsimComparing texts 2023ElabQUDExplaining added details 2024QSalienceWhich questions matter? 2026MQUDQuestions from figures 2024Bilingual disfluenciesPredictability in speech 2026ContextWeaverMemory for agents

Rivers group research themes.→Arrows mark connections between papers.

2023 · Findings of ACL

QUD parsing

Project ↗paper ↗

A parser connects each sentence to an earlier sentence through the implicit question it answers.

Connection to QUDeval QUDeval develops a human evaluation protocol for QUD parsing and examines the questions produced by parsing systems. Abstract ↗

Connection to QSalience QUD parsing recovers the questions a text answers. QSalience asks which questions would most help a reader understand that text. Section 2 ↗

Connection to ContextWeaver QUD parsing connects each sentence to an earlier sentence through the implicit question it answers. ContextWeaver links each reasoning step of an agent to the earlier steps it depends on. Section 3 ↗

2023 · EMNLP

QUDeval

Project ↗paper ↗

Four linguistic criteria test whether a generated question fits its anchor, answer, and preceding text.

Connection to QUD dependency parsing QUDeval develops a human evaluation protocol for QUD parsing and examines the questions produced by parsing systems. Abstract ↗

Connection to QUDsim QUDsim cites QUDeval in its question generation procedure, then uses answerability across documents to measure discourse similarity. Section 4 ↗

Connection to QSalience QSalience follows QUDeval’s constraint that a question must be grounded in its anchor sentence when judging question validity. Section 3 ↗

2025 · COLM

QUDsim

Project ↗paper ↗

The questions answered by two documents reveal similarities in how they develop their ideas.

Connection to QUDeval QUDsim cites QUDeval in its question generation procedure, then uses answerability across documents to measure discourse similarity. Section 4 ↗

2023 · EMNLP

ElabQUD

Project ↗paper ↗

Implicit questions describe the explanations editors add when simplifying an article.

Connection to QSalience QSalience cites ElabQUD to motivate choosing which concepts need an explanation, beyond generating possible elaborations. Section 2 ↗

2024 · EMNLP

QSalience

Outstanding Paper Award

Project ↗paper ↗

We predict how much answering a question would help a reader understand the preceding text.

Connection to QUD dependency parsing QUD parsing recovers the questions a text answers. QSalience asks which questions would most help a reader understand that text. Section 2 ↗

Connection to QUDeval QSalience follows QUDeval’s constraint that a question must be grounded in its anchor sentence when judging question validity. Section 3 ↗

Connection to ElabQUD QSalience cites ElabQUD to motivate choosing which concepts need an explanation, beyond generating possible elaborations. Section 2 ↗

Connection to MQUD MQUD draws on QSalience when extending question salience from news text to scientific figures and a paper’s research goals. Section 2 ↗

2026 · Under submission

MQUD

Project ↗paper ↗

Scientific figures raise questions about a paper’s argument. MQUD studies those questions with annotations from the papers’ authors.

Connection to QSalience MQUD draws on QSalience when extending question salience from news text to scientific figures and a paper’s research goals. Section 2 ↗

2024 · ICML · LLMs and Cognition Workshop

Bilingual disfluencies

Project ↗paper ↗

Language model surprisal and linguistic features help us study disfluencies in monolingual and bilingual speech.

2026 · Under submission

ContextWeaver

Project ↗paper ↗

Dependencies between an agent’s reasoning steps guide which observations and decisions to retain in context.

Connection to QUD dependency parsing QUD parsing connects each sentence to an earlier sentence through the implicit question it answers. ContextWeaver links each reasoning step of an agent to the earlier steps it depends on. Section 3 ↗

Links

Department of Electrical and Computer Engineering

The University of Texas at Austin

The University of Tokyo

---