NYCU NLP LabNational Yang Ming Chiao Tung University

NYCU NLP Lab · Dept. of Computer Science

AI that reasons, adapts, and works with people.

We build personalized, trustworthy, and efficient AI that brings together multimodal perception, knowledge, and reasoning to support people in daily life.

Explore featured research Publications

Peer-reviewed publications

Distinct venues, incl. EMNLP, EACL, SIGIR, WWW

Papers at EMNLP & EACL 2026

Scroll

Research areas

Four questions we keep coming back to.

Four directions, one goal: AI that understands people and the world they live in, and helps them well.

Flagship theme →

Multimodal understanding & situated intelligence

How can AI integrate what it sees, hears, and reads to understand an environment, an event, and what the user needs right now?

EgoArgusVISTA

See papers

Personalization & adaptive intelligence

How can AI use personal knowledge and interaction history to adapt to each user and each task?

PersonaEditRSVPE-QGenISSR

See papers

Trustworthy reasoning & assistance

How can AI reason from evidence and give help that is understandable, checkable, and appropriate?

RefuteClaimMAGICMathEDUFinancial evidence recommendation

See papers

Efficient & resource-aware AI

How can AI allocate models and data to what a task needs, balancing quality against cost?

VDAR-RouterConfidence-driven selectionRepresentative sample selection

See papers

Featured research

Play with the idea.

The question, the mechanism, and what we found.

EMNLP 2026 · Main conference

EgoArgus: Benchmarking VLMs as Situational Assistants

An assistant sees your kitchen and hears you say "the stove is off." Which one should it believe?

Existing egocentric benchmarks test visual understanding in isolation. EgoArgus asks a harder question: when the camera and the user disagree, can the model tell which modality is trustworthy and decide whether to step in at all?

DatasetHuman-annotated dialogue-video scenarios across five daily settings, with visual and language cues that are helpful, irrelevant, or conflicting.

TasksUnderstanding (what is going on) and decision (should the assistant intervene, and how).

FindingCurrent VLMs still struggle to arbitrate between modalities, and existing bias-mitigation methods help only within narrow limits.

Read the paper VISTA companion demo

EMNLP 2026 · Findings

VDAR-Router: Adaptive LLM Routing via Verbalized Query Difficulty

Two queries can use the same words and still be worlds apart in difficulty. Why route them the same way?

Most routers judge a query by its surface semantics or embedding. VDAR-Router first writes out an explicit difficulty analysis, retrieves past queries with a similar difficulty profile, and uses their outcomes to pick the model that balances cost and quality.

Step 1Verbalize the difficulty of the incoming query.

Step 2Retrieve historical records with similar difficulty profiles.

Step 3Estimate each candidate model's suitability and select under a cost-performance reward.

Read the paper Related: confidence-driven selection

WWW 2024

How We Refute Claims: Fact-Checking through Flaw Identification

"False" is not an explanation. What exactly is wrong with the claim?

RefuteClaim reframes fact-checking as flaw-oriented verification: generate the aspects a claim should be judged on, identify the flaw under each aspect, and explain the verdict. FlawCheck, our dataset, distills expert fact-checker reviews into aspects and flaws.

TaskAspect generation and flaw identification, in addition to the verdict.

DatasetFlawCheck, built from expert reviews.

FindingRefuteClaim is especially strong at classifying and elucidating false claims.

Read the paper Watch the demo

EMNLP 2026 · Findings

PersonaEdit: Representative Sample Selection for Personalized Model Editing

A user has told the model hundreds of things about themselves. Which few should be written into its weights?

Retrieval-based personalization depends on retrieval quality and degrades over long-term interactions. Model editing writes new knowledge directly into the parameters, but editing every user sample is costly and the edits interfere with each other. PersonaEdit clusters the hidden representations of a user's data and picks representative samples by proportional stratified sampling.

ProblemScaling model editing to personalization: many edits cost more and interfere.

IdeaCluster hidden representations, then sample representatives in proportion to each cluster's size.

FindingModel editing is an effective route to personalization when the editing samples are chosen well.

Read the paper

EACL 2026 · Main conference

MathEDU: Feedback Generation on Problem-Solving Processes

A student's answer is wrong. Where did the reasoning go off track, and what should the teacher say?

MathEDU pairs real student problem-solving processes with teacher-written feedback, and evaluates models on three hierarchical tasks: is the answer correct, which step is wrong, and what feedback would help.

Task 1Answer correctness classification.

Task 2Error identification at the step level.

Task 3Feedback generation, compared against teacher-written feedback.

FindingFine-tuning helps with correctness and locating the wrong step, but generated feedback still falls well short of teachers'.

Read the paper Position paper (EMNLP 2023)

Flagship theme · Situational assistants

Personal AI assistants that see, listen, and know when to act.

Our largest research line builds assistants that see the world from your point of view, listen to what you say, and decide whether, when, and how to help.

Full series →Overview, timeline, dataset, platform, papers

What the user says

Reviewed decision

Why

Dataset & code EgoArgus paper

Generated by VISTA

What VISTA does

Rare or risky assistance situations are hard, costly, or unsafe to film. VISTA takes a one-line scenario seed and produces an editable plan, an egocentric video, and an auditable review trail. The platform ships 14 reviewed demonstrations across reactive, explicit proactive, and implicit proactive help, in safety-critical and everyday settings, plus matched cases where the right answer is to stay quiet.

This clip, from seed to review

SeedA reflective metal lunch container is held at an open microwave while the user expresses uncertainty.

PlanReveal the container’s material. Show the microwave door opening. Intervene before heating begins.

ExpectedWarn before the container is heated and suggest a microwave-safe dish.

ReviewVisible cues, dialogue cues, why assistance is warranted, and a stated limitation.

Open the VISTA platform VISTA paper

Videos, dialogue, reviewed decisions, intervention times, and assistant messages are taken verbatim from the EgoArgus decision annotations (dataset/paper_eval.csv). The videos were generated with VISTA.

Publications

By year, since 2022.

Systems & demos

Research you can watch running.

Demo systems built by lab members, with short videos.

ACML 2025

ISSR

LLMs for vocabulary test distractor generation with iterative selection and self-review.

IJCAI 2024 Demo

E-QGen

Educational lecture abstract-based question generation.

LREC-COLING 2024

MAGIC

Multi-argument generation with self-refinement for domain generalization in fact-checking.

WWW 2024

RefuteClaim

Automatic fact-checking through flaw identification and explanation.

EMNLP 2023

MathEDU

LLMs and the future of mathematics education.

2024 法律 × 法遵科技黑客松 · 理律學堂跨領域特別獎

金融消費者權益領航員

NLPFIN × NYCU-NLP: refining financial consumer complaints.

People

NYCU NLP Lab

An-Zi Yen 顏安孜

Principal Investigator · Assistant Professor, Department of Computer Science, National Yang Ming Chiao Tung University

Dr. Yen's research interests include natural language processing and information retrieval. Her work has been published in AAAI, EMNLP, WWW, SIGIR, and CIKM. She received the Honorable Mention of the ACLCLP Doctoral Dissertation Award in 2020 and was a distinguished PC member at IJCAI 2023.

Personal page Google Scholar GitHub YouTube

Where

Department of Computer Science, National Yang Ming Chiao Tung University, Hsinchu, Taiwan

What we do

Natural language processing and information retrieval
Personal AI assistants that see, listen, and know when to act

Join us

Master's and PhD students: read a recent paper from the area you like, then write to us with what you would do next.

PhD students

Yu-Chien TangYen-Shyun Huang

Master's students

You-Mei HuangHsiang-Yu ChengDing-Xin KangNien-Hui HuangYong-Xuan ZhengWei-Lin WenTing-Shian KuoLai-Sen CaoHong-Yuan ChengYi-Fu ChenWei-Chia ChangMin-Quan WangTun-Chieh YangChun-Ting Hou

Alumni

Yu-Hang DengBo-Cheng PanMin-Kuang TsuYu-Tung ChouChun-Wei KangGuan-Yi HeWei-Ling HsuYu-Cheng LiuBo-Wei ChenCheng-Xin SongHsiu-Hung LeeYen-Chun HuangYu-Chen KangJou-Chia Chiang

Contact

Interested in joining the lab or collaborating?
azyen@nycu.edu.tw

Prospective students: read a recent paper from the area you like, then write to us with what you would do next.