WING @ NUS | Web IR NLP Group @ NUS

LLMs Infer Cultural Context but Fail to Apply It When Responding

Sun, 25 Oct 2026 00:00:00 +0000

AnalysisBank: An Expert Analysis Pattern Library for Financial Report Generation

Tue, 01 Sep 2026 00:00:00 +0000

Bridging Talk and Thought: Understanding Dialogue Dynamics Across Collaborative Problem-Solving Contexts

Sat, 01 Aug 2026 00:00:00 +0000

REPA: Reproducibility Evaluation via an Autonomous Pipeline Architecture

Tue, 07 Jul 2026 00:00:00 +0000

Yisong Miao Presents His Vector Institute Research Internship Work (with UBC NLP) at the NUS SRI Event

Thu, 02 Jul 2026 00:00:00 +0000

On the afternoon of 2 July 2026, NUS SoC held the 2026 edition of its Summer Research Internship (SRI) Research Gathering and Networking Event. SRI is our school’s program for attracting talent from around the globe to conduct research at our school over the summer. Our group is delighted to host Mr. Thura Aung (KMITL) and Mr. Ha Minh Hieu (HUST) this summer. Dr. Anand Bhojan delivered the opening remarks, and Prof. Chan Mun Choon gave a keynote about SoC’s PhD program.

Yisong Miao, a PhD student from our group, was invited to give a 20-minute presentation at the gathering. He shared his experience as a Vector Institute research intern in 2025, under the supervision of Prof. Vered Shwartz and Prof. Jian Zhu (both from the NLP Group at the University of British Columbia). This is the most interdisciplinary research Yisong has done, connecting pragmatic speaker models (closely related to his PhD topic of discourse) with LLMs’ cultural competence, language grounding, and vision language models. The project examines whether LLMs can act as pragmatic speakers by adapting their responses to users’ cultural backgrounds in conversations grounded in images. The team built Cultural and Pragmatic Response Inference (CAPRI), a new dataset in which users provide cultural cues and ask questions requiring culturally appropriate answers. For example, “What is the temperature?” should be answered in °C or °F depending on the user’s cultural background. Extensive evaluations show that state-of-the-art models can largely infer the user’s cultural background, but often fail to apply this knowledge in their responses. Explicit reasoning about the user’s background, step by step, improves their responses, but can be costly in real-time conversation. The work will appear at EMNLP 2026.

Yisong felt especially honored to give the première présentation of this work at this event, as SRI brings together junior students from around the globe, including Brazil, France, and the USA (which are also represented in the CAPRI dataset!). The seminar concluded with a discussion about future work on extending the CAPRI dataset to more cultures in Southeast Asia (where many seminar audiences come from) and elsewhere around the world.

Prof. Vered Shwartz said: “I’m really excited about this benchmark. Most of the existing work in cultural competence focuses on testing LLMs on their cultural knowledge. Our benchmark allows us to test whether LLMs can also recognize cultural context and apply this knowledge in context. I hope our work will inform future model design. At a personal level, it was a pleasure working with Yisong and Jian on this project. Yisong was always so methodological, organized, and energetic, even though our virtual meetings were happening at 1 am Singapore time. I hope to meet him in person soon!”

Links: Earlier news featuring Yisong as a Vector Institute research intern; The project indexed at our website.

Below is a gallery of the seminar photo

Yisong with the 2026 cohort of NUS SRI interns after the seminar:

Beyond the Crowd: LLM-Augmented Community Notes for Governing Health Misinformation

Wed, 01 Jul 2026 00:00:00 +0000

ComicVQA: A Benchmark for Visual Reasoning in Multimodal LLMs

Wed, 01 Jul 2026 00:00:00 +0000

What's Left Unsaid? Detecting and Correcting Misleading Omissions in Multimodal News Previews

Wed, 01 Jul 2026 00:00:00 +0000

When Does Mixing Help? Analyzing Query Embedding Interpolation in Multilingual Dense Retrieval

Wed, 01 Jul 2026 00:00:00 +0000

CIVIC-AI: Stanford's SALT Lab and WING.NUS Collaborate on Social Intelligence and the Future of AI and Work

Mon, 22 Jun 2026 00:00:00 +0000

The inaugural CIVIC-AI workshop (13–14 July 2026) kicks off a major collaboration between Stanford’s SALT Lab and WING.NUS, bringing together researchers, policymakers, and funding agencies to examine the Social Intelligence of Foundation Models and the Future of AI and Work.

The CIVIC-AI Workshop is an upcoming two-day event in Singapore that will explore the profound societal impacts of AI adoption. The workshop is generously funded by the Ministry of Digital Development and Information of the Republic of Singapore under its AI Visiting Professorship grant, and is also supported by the Ministry of Manpower.

The event centers on three core questions:

Cross-cultural Dimensions: How do cultural context, economic development, and policy environments shape AI adoption and the lived experience of automation across different societies?
Future of Work & Automation: Which professions and tasks are being automated away versus augmented, and who bears the costs of displacement?
Human Advantage: What constitutes irreducible human value when a large fraction of knowledge work can be automated, and how do we measure and support it?

The first day (13 July) focuses on the theme of Social Intelligence, featuring invited talks by researchers such as Yohan Jo, Nancy F. Chen, and Lizi Liao, alongside lightning talks and a panel discussion. The highlight of the day will be a keynote by Diyi Yang from Stanford’s SALT Lab on “Optimizing Human-AI Collaboration,” which will introduce General User Models (GUMs) and Next Action Prediction (NAP) to foster proactive and meaningful human-AI interaction.

The second day (14 July) will center on the Future of Work With AI Agents in Singapore, featuring presentations by Diyi Yang and Min-Yen Kan, followed by extensive discussion sessions to finalize directions and next steps for the collaboration.

Registration for the workshop closes on 6 July 2026. For more details and to register, visit the CIVIC-AI workshop website.

The workshop is made possible by the Singapore Global AI Visiting Professorship (AIVP), an initiative by the Ministry of Digital Development and Information to strengthen Singapore’s AI research capabilities by attracting world-class expertise. Min-Yen Kan (WING.NUS) and Diyi Yang (SALT Lab, Stanford University) are among six NUS Computing faculty teams awarded AIVP grants, for their project “Evaluating and Building Socially Intelligent Foundation Models” — the research collaboration that underpins CIVIC-AI.

FlowSteer: Prompt-Only Workflow Steering Exposes Planning-Time Vulnerabilities in Multi-Agent LLM Systems

Tue, 12 May 2026 00:00:00 +0000

Yuxi Xie Defends Her Ph.D. Thesis on Closed-Loop Scaling

Fri, 08 May 2026 00:00:00 +0000

Congratulations to Yuxi Xie on successfully defending her Ph.D. thesis, “Closed Loop Scaling: Autonomous Improvement of LLM and VLM Reasoning”, from WING!

We wish her every success as she moves into her next chapter.

Many thanks to her committee members, Zhiyong Huang, Anthony K. H. Tung, and See-Kiong Ng, and to her many collaborators, including Michael Shieh, Junxian He, and Kenji Kawaguchi.

Special thanks as well to Prof. William Wang for hosting Yuxi at the University of California, Santa Barbara.

Lyle Ungar Visits WING to Give Invited NAII Seminar on Evaluating LLMs for Cross-Cultural Coaching and Therapy

Thu, 30 Apr 2026 11:00:00 +0800

Did you miss Lyle Ungar’s NUS AI Institute seminar on evaluating large language models for cross-cultural coaching and therapy?

Watch the recording to learn how frontier LLMs still struggle with culturally appropriate conversation and what can be done about it.

Lyle Ungar (University of Pennsylvania) explores how to evaluate LLMs on subjective tasks like coaching and therapy across cultures. The talk highlights:

An interpretable framework for comparing linguistic politeness cues across languages
Evidence that LLMs can homogenize non-Western styles toward stylistic neutrality, a failure missed by standard metrics
RASTA, a retrieval-based method that preserves culturally appropriate style without hurting content
The Culturally-Aware Conversations benchmark

Lyle Ungar is Professor of Computer and Information Science at the University of Pennsylvania, with cross-appointments in Psychology, Bioengineering, Genomics & Computational Biology, and Operations, Information & Decisions. His group builds explainable NLP and socio-emotionally sensitive AI for psychological and medical research.

This NUS AI Institute seminar was recorded and rebroadcast with explicit permission from Lyle Ungar.

Recording: Watch the seminar

Improving Conversational Recommendation with Contextual Adaptation of External Recommenders and LLM-Based Reranking

Mon, 30 Mar 2026 00:00:00 +0000

HistoricRAG: Evidence-Centered Newspaper Retrieval for Misinformation-Resilient Question Answering

Tue, 24 Mar 2026 00:00:00 +0000

Conversing with the Past: Immersive Access to Cultural Digital Archives with LLMs and Spatial Interfaces

Mon, 23 Mar 2026 00:00:00 +0000

WING Receives Singapore AIVP Funding for Socially Intelligent Foundation Models

Thu, 19 Mar 2026 00:00:00 +0000

Big news from WING: our group at NUS Computing, in collaboration with Prof. Diyi Yang at Stanford, has received funding support under Singapore’s AI Visiting Professorship (AIVP).

The grant supports our collaboration with Prof. Diyi Yang on socially intelligent foundation models, including work that advances their cognitive, situational, and behavioral intelligence.

As AI systems become increasingly integrated into human life, there is a growing need to evaluate and build socially intelligent foundation models that are socially aware and human-centered.

Funded by the Ministry of Digital Development and Information, the project, “Evaluating and Building Socially Intelligent Foundation Models,” works toward more trustworthy, human-centered AI. We are excited about what comes next.

Read more: five other professors from the NUS School of Computing were also awarded AIVP grants in the same round, six in all.

Quantitative Computational Validation of Nanoscale Interactions between Drug Molecules and Diabetic Wound-Related Proteins for Drug Discovery

Mon, 02 Mar 2026 00:00:00 +0000

Philipp Mayr Visits WING: AI Overload and Scholarly Document Processing

Fri, 20 Feb 2026 00:00:00 +0000

We were delighted to host Philipp Mayr (GESIS - Leibniz Institute for the Social Sciences) at our group meeting on February 20, 2026. He introduced the Knowledge Technologies department, the Information and Data Retrieval team, the OMINO project, and two Scholarly Document Processing projects.

One highlight was the recent paper “AI Overload: A Multi-Level Taxonomy and the Path Forward” (IEEE Intelligent Systems, 2026), which defines AI overload as a sustained mismatch between AI-amplified demands and human or institutional capacity for supervision and validation.

The talk discussed AI overload across individuals (cognitive strain), organizations (coordination burden), and societies (polarization), and examined overload patterns such as cognitive, informational, coordination, monitoring/control, normative, and affective burdens.

A key insight was that in modern AI systems, the bottleneck is no longer generation, but oversight. As AI autonomy grows, validation, monitoring, and accountability demands increase. Mitigation therefore requires bounded autonomy, calibrated reliance, governance mechanisms, and stronger AI literacy.

Paper: AI Overload: A Multi-Level Taxonomy and the Path Forward

We thank Philipp for the inspiring and thought-provoking talk!

Below is a gallery of the seminar, photo credit to Sahaj!

GitSearch: Enhancing Community Notes Generation with Gap-Informed Targeted Search

Mon, 09 Feb 2026 00:00:00 +0000

Open-source AI for science: OpenScholar and why replicability matters

Wed, 04 Feb 2026 00:00:00 +0000

Min was quoted in Science on the value of open, replicable AI for research: It’s very important to put this type of open-source research out there because it is replicable."

A Science article reports on OpenScholar, an AI built by the Allen Institute for AI (Ai2) and five universities to help researchers keep up with the flood of papers (over 4 million in 2024). Unlike general-purpose chatbots, it answers science questions by searching a database of 45 million open-access papers, synthesising multiple sources, and critiquing its own answers to cut down on hallucinated references. In a Nature study, it outperformed GPT-4o and Llama on domain benchmarks and was preferred over human experts in many cases. The team have made the code and data freely available—which Min highlights as crucial for peer review and replicability. The piece also notes caveats: paywalled literature is missing from the corpus, and over-relying on summaries can deskill researchers and obscure nuance. A follow-up model, DR Tulu-8B, was described in a November 2025 preprint.

Epistemic Context Learning: Building Trust the Right Way in LLM-Based Multi-Agent Systems

Fri, 30 Jan 2026 00:00:00 +0000

Michalis Vlachos Presents at the CSSH Seminar: Searching Digital Libraries with Language Models

Wed, 28 Jan 2026 00:00:00 +0000

Our group is delighted to have Michalis Vlachos visiting our group as a Sabbatical Visiting Professor. On January 28th, 2026, Michalis presented his work on “Searching Digital Libraries with Language Models” at a seminar hosted by the Centre for Computational Social Science and Humanities (CSSH) at NUS.

Abstract:

Digital libraries now hold tens of millions of pages, yet most are still accessed through keyword based search over potentially noisy OCR text. As collections expand, traditional search interfaces struggle to support meaningful discovery, context, and evidence-based answers.

This talk explores how modern language models can enhance the entire digital library pipeline. We examine how LLMs can refine OCR output, clean historical text, and enable natural-language search through retrieval augmented generation. Using real-world archival data, we show improvements in character and word error rates, as well as downstream gains in retrieval quality, evaluated through both standard metrics and LLM based judging for faithfulness, correctness, and relevancy.

We also look beyond ranking results, discussing evidence driven answers and immersive, augmented-reality interfaces that open new ways to explore large historical collections. We conclude by reflecting on how these advances can improve transparency, reduce misinformation, and reshape the future of search in digital libraries.

Below is a gallery of the seminar, photo credit to Min and Yisong!

Min's AAAI workshop keynotes (26–27 Jan 2026)

Tue, 27 Jan 2026 00:00:00 +0000

Min gave two invited keynotes at AAAI 2026 workshops on 26 and 27 January, entitled “The Research Manifold” and “Top 5 trends to watch in the future of IR research (and why they are irrelevant)”. Both talks are now available to watch.

26 Jan — The Research Manifold

Min’s first keynote addressed research principles and open challenges for the community.

Imagine research as a complex, multidimensional vector space where human researchers are partially observable points in this space, with vectors of velocity and acceleration mapping, and where their new discoveries and publications mark observations. This research manifold is now being disrupted by the new entrant of artificially intelligent agents and tools. I describe my thoughts on some of the challenges ahead, as we progress to centaur-like human–AI research teams.

Watch the keynote (workshop page with video).

27 Jan — Top 5 trends to watch in the future of IR research (and why they are irrelevant)

The second keynote highlighted five central themes or open problems in information retrieval.

What are the likely trends in the future of information retrieval for 2026? I discuss five trends, give evidence for their provenance and impact, but you will have heard about many of these trends from previous speakers. Paradoxically, it may be better for some in the field not to follow these trends and I discuss why.

Watch the keynote (YouTube).

For other talks, see other WING conference presentations on our group’s YouTube playlist.

WING 2026 Spring Outing at Clarke Quay

Fri, 23 Jan 2026 00:00:00 +0000

WING members were excited to have a spring outing at Clarke Quay on January 23, 2026. For this outing, we scheduled it at the beginning of the semester to help new members better cohere with the group.

We enjoyed both the beautiful river view and the pizza! We were happy to see the return of our alumni, especially Hengchang, Victor, and Taha, as well as our collaborator, Yunshan. We especially thank Hengchang (a WING PhD alumnus) for his generous donation of the pizza (and other items on the bill as well)!

Below is a gallery of the outing, photo credit to Yajing, Glenda, and Yisong!

WING Hosts a Lab Tour for AAAI 2026 Participants

Fri, 23 Jan 2026 00:00:00 +0000

AAAI 2026 has been held in Singapore from January 20 to January 27. Some of our group members are excited to travel to the conference site (at Expo) to communicate research. In conjunction with our regular group meeting on Friday, January 23, we are delighted to invite AAAI participants to visit our WING lab and experience the NUS Computing environment.

With our visitors, this group meeting had a special agenda:

Minzhi Li from WING NUS gave a presentation to debrief her 2025 internship experience.

Dr. Lu Gan (from GESIS - Leibniz Institute for the Social Sciences) gave a presentation about an AAAI paper from their team, titled “GSAP-ERE: Fine-Grained Scholarly Entity and Relation Extraction Focused on Machine Learning” (available here).

Finally, we had a round of self-introductions for all WING members as well as the visitors to socialize and discuss research.

Below is a gallery of the lab tour, photo credit to Sahaj, Yisong, and Min!

SSNLP 2026: WING Members Present Three Posters

Wed, 21 Jan 2026 00:00:00 +0000

The Singapore Symposium on Natural Language Processing (SSNLP) returns on Wednesday, January 21, 2026, at SUTD as a full-day, in-person event. Since 2018, SSNLP has been the annual gathering of Singapore’s NLP community, bringing together students, faculty, and industry researchers to share ongoing work and spark collaboration. Held in conjunction with AAAI 2026, this year’s symposium will feature invited keynotes from leading researchers in the field, a panel discussion, and an immersive poster-focused program showcasing recent work from Singapore-based researchers published in top venues such as ACL, EMNLP, NeurIPS, ICLR, and AAAI this year (this short intro is propagated from SSNLP 2026 website).

WING members are delighted to attend the seminar, enjoying keynote talks from guest researchers who travel all the way from AoE to Singapore. We also appreciate the panel discussions led by PC Chairs (from *ACL, ICLR, and NeurIPS conference), who share valuable insights on the peer review process, particularly from the students’ perspective.

We also engage in the immersive poster program, and our group is excited to present three posters:

Discursive Circuits: How Do Language Models Understand Discourse Relations? by Yisong Miao and Min-Yen Kan (EMNLP 2025)
SkyLadder: Better and Faster Pretraining via Context Window Scheduling by Tongyao Zhu, Qian Liu, Haonan Wang, Shiqi Chen, Xiangming Gu, Tianyu Pang, and Min-Yen Kan (NeurIPS 2025)
Drifting Away from Truth: GenAI-Driven News Diversity Challenges LVLM-Based Misinformation Detection by Fanxiao Li, Jiaying Wu, Tingchao Fu, Yunyun Dong, Bingbing Song and Wei Zhou (AAAI 2026).

We thank the symposium organizers for the opportunity to present our work, as well as the enthusiastic audience who engaged with us during the poster sessions!

Below is a gallery of the poster presentations, we thank Prof Min for taking these photos!

Yisong presenting his work on “Discursive Circuits: How Do Language Models Understand Discourse Relations?”:

Tongyao presenting his work on “SkyLadder: Better and Faster Pretraining via Context Window Scheduling”:

Fanxiao presenting his work on “Drifting Away from Truth: GenAI-Driven News Diversity Challenges LVLM-Based Misinformation Detection”:

The Facade of Truth: Uncovering and Mitigating LLM Susceptibility to Deceptive Evidence

Fri, 09 Jan 2026 00:00:00 +0000

WING is looking for postdoctoral scholars and doctoral students!

Mon, 15 Dec 2025 00:00:00 +0000

The Web, Information Retrieval, and Natural Language Processing (WING) group at the School of Computing, National University of Singapore is actively seeking postdoctoral scholars and doctoral students to start in 2026 across three cutting-edge research areas: AI for Science, LLM Socioethical Alignment, and Fact Checking. WING is led by Associate Professor Kan Min-Yen.

Postdoctoral and Doctoral Student Positions Available at WING Lab, NUS (2026)

Research Areas

AI for Science

We are seeking postdocs interested in cross-disciplinary research, particularly in computational social science and humanities (NUS CSSH). Scholars will collaborate with experts specializing in Southeast Asian studies, public policy, materials science, and computer science topics, fostering innovative approaches to scientific discovery through AI.

LLM Socioethical Alignment

This position involves working with a cross-institutional team spanning NUS and US institutions on cultural and safety alignment of Vision Language Models (VLMs). The project aims to make VLMs more culturally sensitive and personalized for deployment in the multicultural and multilingual environments common throughout Southeast Asia.

Fact Checking

As part of the NUS Centre for Trusted Internet and Community (CTIC), scholars will conduct research to help both readers and producers of social media better understand and detect misinformation. The focus is on distinguishing nuance and developing improved information literacy practices.

Postdoctoral Position Qualifications

Successful applicants should demonstrate:

Deep and rigorous research training demonstrated by a doctoral degree in computer science or related area from an advising group well-known for natural language processing, machine learning, digital libraries, and information retrieval research
Strong record of research, often demonstrated by research publications at premier venues, including ACL, SIGIR, ICML, and ICLR
Two or more reference letters from current supervisors or employers

Shortlisted postdoctoral applicants will participate in a comprehensive evaluation process including a job talk, technical interview, a criticial writing assignment, and a interview with the lab head.

Position Details

Location: On-site at the School of Computing, National University of Singapore
Start Date: As soon as possible in 2026
Contract: 3-month probationary period, followed by 2-year contract

Career Development: Mentoring aligned with individual career goals, including opportunities for:

Undergraduate, Masters, and PhD student supervision
Grant writing and proposal development
Classroom teaching experience
Research collaboration and publication

Doctoral Student Qualifications

Deep and rigorous research training demonstrated by a doctoral degree in computer science or related area from an advising group well-known for natural language processing, machine learning, digital libraries, and information retrieval research
Strong record of research, often demonstrated by research publications at international venues

Shortlisted doctoral student applicants will need to pass a criticial writing assignment, a software test and an interview with the lab head. Doctoral applicants must also meet the rigourous admission standards of the School of Computing. If you haven’t yet applied to be part of SoC, we hope that you will consider SoC for your graduate studies, irrespective of whether WING can host your doctoral studies or not. If you are admitted, you can try to contact us again after you arrive in Singapore. If you have questions about graduate studies at SoC NUS, your best bet is to ask our very knowledgeable staff at the graduate office (phd-info@comp.nus.edu.sg). They know the ins and outs of the application process, deadlines and requirements.

Application Process

To apply, please send an email to knmnyn@nus.edu.sg, including “WING” in the subject line, clearly specifying which topic(s) you are applying for.

For postdoctoral candidates, reference letters should be sent directly to the same email address by your referees under separate cover.

Due to the high volume of applications, only shortlisted candidates will be contacted. For more information about WING’s research activities and recent publications, visit our lab’s website.

WING places an emphasis on the diversity of our Ph.D. cohort so that the group can benefit from a wider perspective, so WING generally limits the number of students and staff from well-served demographics (e.g., male Chinese students).

ChatGPT 5.2 and Nano Banana was used to create the image. Claude 4 was used to refine the advertisement text.

Jiaying represents iGyro project at NeurIPS 2025 in San Diego

Wed, 10 Dec 2025 00:00:00 +0000

Jiaying helped to represent WING’s interest in the iGyro information resilience project at the NUS Computing networking session recently at NeurIPS 2025.

Jiaying writes: Grateful for the opportunity to represent the NUS Centre for Trusted Internet and Community (CTIC) at the first NUS Computing Network Session and Reception during #NeurIPS2025! It was wonderful engaging with brilliant attendees from around the world and sharing CTIC’s research highlights on digital resilience, alongside our amazing NUS Computing faculty and researchers. Kudos to the organizers for putting together such a meaningful and energizing event! 🥳

Bryan Low (the organiser of the NeurIPS 2025 networking session for our School of Computing) also notes the unique “3C"s attributes of our school:

(1) We provide environment & resources to champion bold initiatives & tackle grand challenges. As a result, we can go beyond publications to produce research breakthroughs, societal impact & thought leadership.

(2) We nurture, mentor & cultivate the young generation to be future leaders who will decide the society of the future.

(3) We embrace the kampung spirit that is uniquely Singapore. Imagine a melting pot of diverse global talents coming together to form a community/kampung that is collegial, collaborative, and caring.

Yisong featured as a Vector Institute Intern

Mon, 08 Dec 2025 00:00:00 +0000

Check out 👀 Yisong Miao’s feature as a Vector Institute research intern in the latest 2025 installment of their news cast. Although he wasn’t able to travel to the institute in 🇨🇦 for the immersive physical part, WING is proud that he was able to participate in the programme while staying on the other side of the world!

Yisong mentored with Vector’s Vered Shwartz to understand why LLMs assume certain contexts and what helps them adapt to different cultures in what was his “most interdisciplinary project ever.” As a Vector intern, Yisong also got to network with IMPACT mentors. We wish him continued success in his research work as he continues his work at WING.

see the original LinkedIn post

MuSLR: Multimodal Symbolic Logical Reasoning

Mon, 01 Dec 2025 00:00:00 +0000

Self-Improving and Self-Adapting Agents

Mon, 01 Dec 2025 00:00:00 +0000

🎯 Overview: We study test-time scaling (TTS) methods for self-improving and self-adapting agents, advancing a new paradigm of artificial intelligence in which autonomous systems do not merely act, but evolve reliably through learning from experience, refining behavior at test time, and autonomously modifying their own learning mechanisms.

🧠 Abstract: Although modern foundation models (FMs) like ChatGPT are extraordinarily capable, they remain largely fragile: small variations in input (aka ‘‘prompts’’), such as subtle phrasing changes or unintended noise, can lead to contradictory or ungrounded reasoning. This limits their broader deployment in high-stakes domains like healthcare, law, and science, where precision, reliability, and interpretability are essential.

In this project, we aim to build the next generation of FM systems that can continually assess and correct their behavior at test time, making them more trustworthy and capable. Our studies span three main aspects:

(1) Understand Fragility of FMs: How can we interpretably understand model behavior through the lens of prompts to improve FMs reliably? (NLPromptEval @ ACL'25, FormatBiasEval @ NAACL'25)
(2) Study TTS Methods: With the above understanding, how can we develop TTS to enhance model effectively and reliably? (Multi-Expert Prompting @ EMNLP'24, adv-ICL @ ACL'24, Chain-of-Opinion @ COLING'25)
(3) Self-Improving Agents: How can we build powerful TTS-powered agentic systems that critique and refine their own behavior, ultimately enabling autonomous self-improvement? (LongGuide (Self-Adapting Agent) @ ACL'25, VISTA (Self-Improving Agent))

Our findings so far have offered a foundation for building foundation model agents that reliably improve through interaction, feedback, and structure-aware test-time methods—without reliance on additional gradient-based training.

🔥 News: We are expanding our studies towards multiple directions. Please review our work above. If you have a strong interest in the topics listed and a solid background in mathematics and programming, and optionally in research, come to join us!

SkyLadder: Better and Faster Pretraining via Context Window Scheduling

Mon, 01 Dec 2025 00:00:00 +0000

WING Members Present at NUS AI Seminar at the Computer Science Department

Wed, 19 Nov 2025 00:00:00 +0000

NUS AI Seminar at the Computer Science Department is a new initiative to bring together AI researchers to share their latest research findings and insights. There are eight speakers in the Fall 2025 semester, usually hosted on Wednesday afternoons at Seminar Room 21. WING members are excited to present their research alongside other presenters from NUS Computing, MIT, and UCLA. We appreciate Prof. Lee Wee Sun and PhD student Guoji Fu for hosting the seminar.

Yisong Miao gave a seminar presentation titled “Faithfulness and Interpretability for Discourse Understanding” on November 19th, 2025. This talk combines his work on Discursive Socratic Questioning (ACL 2024) and Discursive Circuits (EMNLP 2025). He explained how our team tackles the technical challenge of evaluating a model’s discourse faithfulness and formalizing a unique setup for transformer circuit discovery. The seminar ends with an audience discussion on the conceptual difference between truthfulness and faithfulness, as well as designing new methods using sparse circuits in future work.

Yajing Yang presented “From Data to Insights: LLMs for Financial Narration”, also on November 19th, 2025, highlighting her research on financial data narration. Drawing from DataTales: A Benchmark for Real-World Intelligent Data Narration (EMNLP 2024) and KAHAN: Knowledge-Augmented Hierarchical Analysis and Narration for Financial Data Narration (EMNLP 2025), she described our team’s efforts to build a realistic benchmark and improve generation quality through knowledge integration and hierarchical analysis. The talk sparked discussion on remaining challenges in financial report generation and how LLMs could be used to produce more actionable and informative insights.

On December 3rd, 2025, Xuan Long Do delivered a seminar titled “Toward Autonomous Self-Improving Foundation Models”. The presentation spanned multiple projects, including “What Makes a Good Natural Language Prompt?” (ACL 2025), Multi-Expert Prompting (EMNLP 2024), and VISTA, his 2025 Google internship project on a self-improving video generation agent. Beginning with an analysis of prompt and model fragility, he showed how multi-agent reasoning can help address these limitations before moving toward the broader vision of autonomous self-improving foundation models. The seminar concluded with an audience discussion on future research directions in this area.

Below is a gallery of the seminar photos, some of which are enhanced and regenerated via Google’s Nano Banana

Yisong Miao presenting his work on “Faithfulness and Interpretability for Discourse Understanding” at the AI Seminar on November 19th, 2025:

Yajing Yang presenting her work on “From Data to Insights: LLMs for Financial Narration” at the AI Seminar on November 19th, 2025:

Xuan Long Do presenting his work on “Toward Autonomous Self-Improving Foundation Models” at the AI Seminar on December 3rd, 2025:

WING Members Present at NUS Computing's Inaugural AI-X: Industry Day

Tue, 18 Nov 2025 00:00:00 +0000

A gallery of additional photos from the event is included below, with thanks to the official event photographer.

WING Members Enjoy the Group Outing at the Macritchie Reservoir

Fri, 14 Nov 2025 00:00:00 +0000

A gallery of additional photos of the outing is included below.

WING's CS6101 Retrieval Augmented Generation comes to a successful close at 27th STePS

Wed, 12 Nov 2025 00:00:00 +0000

WING ran an experimental section of our CS6101 Exploration of Computer Science Research, as a trial course on Retrieval Augmented Generation.

RAG is now a common gateway experience for the general public as well as students to experience Information Retrieval, a core research area of WING.

The CS6101 course website has publicly-available lecture slides, video lecture recordings and scribe notes for the class’s topics.

These include sections on LLM generation, Retreival, Ranking, RAG Models, Iterative Retrieval and Advanced RAG topics (inclusive of graph, multimodal RAG) and source credibility estimation.

Check out the LinkedIn posts by class participants Vangmay Sachin, Jonathan Chen and Sahej Agrawal.

You can also visit the other WING Events, including past public iterations of WING’s CS6101 sections.

Image enhanced and regenerated via Google’s Nano Banana

KAHAN: Knowledge-Augmented Hierarchical Analysis and Narration for Financial Data Narration

Wed, 05 Nov 2025 00:00:00 +0000

Discourse and LLMs

Tue, 04 Nov 2025 00:00:00 +0000

Large language models (LLMs) can now generate coherent text, but can they truly understand language? In this project, we explore discourse understanding, a foundational objective in natural language processing (NLP). In the first part, we refine the notion of understanding as faithful behavior under a series of Socratic-style questions (appeared at ACL 2024). In the second part, we take a causal approach to mechanistically interpret the key components of transformer models in discourse processing (appeared at EMNLP 2025).

Yisong, a PhD student at WING, began his research journey by tackling the fundamental question: Do language models truly understand text? Over time, he refined this broad inquiry into the specific challenge of discourse understanding. In his recent work, advised by Prof. Min, they introduced a more precise framework for measuring discourse comprehension from three discourse-specific perspectives, while also contributing insights into broader LLM challenges such as logical consistency. Currently, Yisong and Min are delving deeper into the mechanisms behind how models process and understand discourse, and they discovered that sparse computational graphs, termed as “discursive circuits”, control how models process discourse relations.

Discursive Circuits: How Do Language Models Understand Discourse Relations?

Tue, 04 Nov 2025 00:00:00 +0000

LitePruner project awarded VIP@SoC 10K development grant

Thu, 25 Sep 2025 00:00:00 +0000

Sahej’s LitePruner wins a Venture Intiation Programme award of 10,000 SGD to prototype and build out a prototype for commercialisation and use in production large language models!

LitePruner compresses LLM prompts to cut costs and latency. It works across APIs without retraining and preserves accuracy. A prototype targets high-volume enterprise users. Sahej is supervised by Dr Barid Xi Ai.

For the VIP, Sahej will be mentored by Prof. Yan Shuicheng.

About the Venture Initiation Programme at the School of Computing:

Is your idea unique or something that has been tried before? (Check out the past projects here for reference)

Is there a competitive or differentiating edge? How is it different from what already exists and how is it better?

Is there a sizable market and a clearly defined initial target user group for your proposed innovation?

Have you done sufficient research on the market need and competition?

Does the team have the competencies to deliver?

Are you able to demonstrate some early prototypes? (This would be an advantage!)

What counts as plagiarism? AI-generated papers pose new risks

Wed, 20 Aug 2025 00:00:00 +0000

Min was quoted in the news feature on Nature Communications: “The important thing is these [AI] tools are going to be here. We need to find the right way of using them”.

In a Nature news feature on “What counts as plagiarism? AI-generated papers pose new risks”, writer Ananya quoted joint research work from WING; our collaboration with Univ. of Siegen’s Dr. Joeran Beel on Sakana’s AI Scientist and its promises, challenges and implications for Information Retrieval research.

Read our original pre-print here: Evaluating Sakana’s AI Scientist for Autonomous Research: Wishful Thinking or an Emerging Reality Towards ‘Artificial Research Intelligence’ (ARI)?.

Image generated by ChatGPT 5.

Fact checked: Dr. Xinyuan Lu graduates!

Thu, 14 Aug 2025 00:00:00 +0000

Congratulations to Dr. Xinyuan Lu on successfully defending her PhD thesis entitled " Fact-Checking Complex Claims using Large Language Models"!

We are thrilled to announce that Dr. Xinyuan Lu has successfully completed her doctoral defense! Her thesis focuses on the complexity in fact-checking systems, especially with respect to claims that require mulit-step and tabular reasoning. Dr. Lu is in the final stages of her job search.

A special thanks to her examiners Profs. See-Kiong Ng (first from the left), Mong-Li Lee (second from left) and Prof. Roger Zimmermann (right).

Learn more about Dr. Lu at her Homepage. You can follow her on X as @ShellyLu00).

CS6101 T2510 - Retrieval Augmented Generation

Thu, 07 Aug 2025 00:00:00 +0000

Beyond In-Context Learning: Aligning Long-form Generation of Large Language Models via Task-Inherent Attribute Guidelines

Fri, 01 Aug 2025 00:00:00 +0000

Project Check Up's 1 year update

Fri, 01 Aug 2025 00:00:00 +0000

Our Crazy Ideas project, CheckUp: Grounding Social Media on Healthcare with Science, has made strong progress in advancing the dissemination of reliable health information on social media.

We introduced HashView, a benchmark dataset for predicting hashtag popularity, and developed BuzzProphet, a reasoning-augmented framework that combines insights from large language models with lightweight regressors. This research has been accepted at CIKM 2025, highlighting the novelty and relevance of our approach. The project has also contributed to capacity building by supporting one Research Assistant and mentoring two students, including one MComp graduate and one ongoing undergraduate FYP student.

Building on CheckUp’s early findings, we have secured a S$ 100K Ministry of Education Tier 1 grant entitled “Automating Community Note Generation for Social Media Misinformation Governance”. The groundwork laid by CheckUp demonstrated how LLMs can enhance the clarity and contextual relevance of social media content, directly informing the design of this new initiative. While CheckUp focuses on proactive communication of credible information, the Tier 1 project addresses the post-exposure stage of misinformation by developing real-time, AI-driven Community Notes as corrective tools. Together, these efforts form a complementary strategy for reducing the societal harms of health misinformation.

Summarizing Scientific Content with Social Media Language

Fri, 01 Aug 2025 00:00:00 +0000

As social media dominates public information consumption, effectively communicating scientific knowledge through these platforms is crucial. Despite advances in document summarization, complex scientific jargon remains a barrier for general audiences, limiting access to important scientific information. This project develops methodologies to simplify scientific content while preserving accuracy, translating it into language that engages social media users. Our goal is to increase accessibility of scientific knowledge and enhance its influence on informed societal decision-making.

What Makes a Good Natural Language Prompt?

Fri, 01 Aug 2025 00:00:00 +0000

Navigating Ethical Challenges in NLP: Hands-on strategies for students and researchers (Tutorial at ACL '25)

Sun, 27 Jul 2025 00:00:00 +0000

Min's IT:U Keynote: The Boy Who Harnessed the Wind: Getting more out of LLM Prompting

Sat, 26 Jul 2025 00:00:00 +0000

Min gave an invited talk at the IT:U Summer School on Natural Language Processing based on Do Xuan Long’s papers over the last three years, based on his papers on prompting.

(26 Jul 2025, Linz, Austria)

What does a famine in Malawi have to do with effective prompting? Like William Kamkwamba’s perseverance in bringing electricity and irrigation to his home, we show how a systematic investigation of prompting mechanisms brings about improved performance across many tasks. This talk describes our work to categorise and structure advice on LLM prompting and investigate how details in prompting can positively impact performance on complex, long generation tasks as well as adherence to output formats. Joint work with NTU and A*STAR, Singapore.

Slides: http://soc-n.us/250726-prompt Video: http://soc-n.us/250726-prompt-yt

Min participates in the AI in Educational panel at the IT:U Austria NLP Summer School

Fri, 25 Jul 2025 00:00:00 +0000

Min gave an opening statement about the importance of socioethical issues in the education process in the age of artificial intelligence.

Here was Min’s opening statement.

I was from a not well-to-do family but an open account left at an internet cafe allowed me access. GPT 7 raised me well, I learned with its guidance and became skilled and respected for commanding AI to better research from a multicultural perspective. I became a student tutor early and continued this into my university education. Sometimes AI gave me feedback sometimes not, but I understood that the "scaffold" was there to help challenge myself, to build myself and resolve; like level design in games — not quite out of reach but certainly a stretch. Now we have our own teaching feedback, our own loss function. My teacher manages me, and I wonder how different it is with me managing my students. I find social justice that being nice and communicating well still serves a good purpose because AI senses that such qualities are helpful in achieving its own goals. I feel a sense of purpose but also a sense that I may have been waylaid into this sense of purpose. Determinism is always at the back of my mind.

Min's invited talk on prompting at the the IT:U Austria Summer School

Fri, 25 Jul 2025 00:00:00 +0000

Check out Min’s talk that makes an analogy to the boy who harnessed the wind with LLM prompting:

Pictures of the NLP Summer School

(Courtesy the IT:U Austria staff. Pictures cleared by the participants with informed consent.)

LLMs Are Biased Towards Output Formats! Systematically Evaluating and Mitigating Output Format Bias of LLMs

Mon, 05 May 2025 00:00:00 +0000

iGyro Symposium on Mis-, Dis-, and Mal-information (iGYRO SMDM 2025)

Mon, 28 Apr 2025 00:00:00 +0000

Reasoning and Planning for Large Language Models

Mon, 28 Apr 2025 00:00:00 +0000

SSNLP '25

Wed, 23 Apr 2025 00:00:00 +0000

Demonstration of Multilingual QA Check to NUS President Tan Eng Chye

Thu, 20 Mar 2025 00:00:00 +0000

NUS President Tan Eng Chye visited the Centre for Trusted Internet and Community, and took part in hands-on demonstrations by members of the iGryo research team. WING-NUS demonstrated our Multilingual QA Check system, which performs Retrieval Augmented Generation and has multilingual capabilities to check claims in local Southeast Asian languages as well as in code-switched (mixed) sentences."

Dr. Victor Li Chuang defends his Ph.D. Thesis!

Thu, 13 Mar 2025 00:00:00 +0000

We are delighted to announce that Dr. Victor Li has successfully defended his PhD thesis titled, “Towards Holistic and Practical Conversational Recommender Systems,” under the co-supervision of A/P Min-Yen Kan and Prof. Haizhou Li, on Thursday, 13 March 2025! He joins the select group of WING.NUS’ doctoral alumni! Congratulations!

Learn more about Dr. Victor Li Chuang at his Homepage. He is currently on the job market!

Misinformation, Disinformation, and Generative AI: Implications for Perception and Policy

Sat, 01 Feb 2025 00:00:00 +0000

A Comparative Analysis of Representation Flow in State-Space and Transformer Architectures