LLM | Web IR NLP Group @ NUS

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”:

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!

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:

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.

ChatCRS: Incorporating External Knowledge and Goal Guidance for LLM-based Conversational Recommender Systems

Wed, 01 Jan 2025 00:00:00 +0000

This project is inspired by the idea of applying LLMs for zero-shot conversational recommendation. Advanced LLMs (e.g., ChatGPT) are limited in domain-specific CRS tasks for 1. generating grounded responses with recommendation- oriented knowledge, or 2. proactively leading the conversations through different dialogue goals. In this work, we first analyze those limitations through a comprehensive evaluation, showing the necessity of external knowledge and goal guidance which contribute sig- nificantly to the recommendation accuracy and language quality. In light of this finding, we propose a novel ChatCRS framework to decompose the complex CRS task into several sub-tasks through the implementation of 1. a knowledge retrieval agent using a tool-augmented approach to reason over external knowledge bases and 2. a goal-planning agent for dialogue goal prediction. Incorporating those external inputs, LLMs proactively plan interactions and generate outputs with rich information. Experiments on two multi-goal CRS datasets reveal that ChatCRS sets new state-of-the-art benchmarks, improving language quality of informativeness by 17% and proactivity by 27%, with tenfold recommendation accuracy enhancement

Model-generated Code-mixed Sentences

Wed, 01 Jan 2025 00:00:00 +0000

Code-switching/code-mixing is a common linguistic practice where speakers switch between multiple languages in a single discourse. Developing language models with code-switching capabilities is crucial to the interest of the large multilingual communities. However, collecting real-world code-mixed sentences for training remains challenging due to its colloquial nature, highlighting the importance of synthetic code-mixed data. Preliminary experiments shows LLMs can generate code-mixed sentences by switching entities in multiple languages. However, these model-generated sentences are often not natural limiting further leverage for training. This project explores methods to improve the naturalness of model-generated code-mixed sentences. Ultimately, we aim to build an automated pipeline capable of generating natural-sounding code-switched sentences for further downstream tasks.

Assessing the AI Scientist

Mon, 26 Aug 2024 00:00:00 +0000

Our collaboration with Joeran Beel and Moritz Baumgart of ISG Siegen originated from Joeran’s sabbatical visit with WING in Oct 2024. Moritz, a talented undergraduate student at University of Siegen, joined us as a research assistant, working remotely which coordinated with his advisor here in NUS.

Over the course of Joeran’s visit, we were able to install and test the AI Scientist system advertised as an automated robo scientist by Sakana.AI. We find positive use cases, but also weaknesses that need much handholding from us.

Haomeng, Rui Feng and Nura join our team as students working towards assessing the growing variety of LLM based research agents going into 2025.

Emoji Composition

Mon, 20 May 2024 00:00:00 +0000

The project began when Yisong, intrigued by the Emoji Mashup Bot, explored the question: Can emojis be composed to convey meanings like English phrases? With the support of his advisor, Prof. Min, he recruited an undergraduate student, Zi Yun Yang, to collaborate on the project. Ms. Yang built an annotation platform to collect emoji compositions and their corresponding English phrases from her peers. After her graduation, Yisong and the team added another layer of annotation on composition types and formalized an emoji-based textual entailment task (EmoTE).

The first iteration of this work was published at LREC-COLING 2024. Prof. Min was not listed as an author due to his role as a conference chair, but this project cannot be finished without his guidance. The project continues to gain interest, particularly among undergraduates, with many students from our local NLP course using the emoji dataset for their final projects.

Visualising Large Language Model Activations

Fri, 17 May 2024 00:00:00 +0000

In the field of natural language processing, large and complex neural networks, particularly transformers, have become essential due to their effectiveness in handling text data. However, despite their success, current methods for visualizing how activations of specific network units influence model outputs remain underdeveloped. This gap hinders interpretability and limits researchers’ and practitioners’ understanding of neural network behaviors during training and inference.

This project aims to address this issue by creating a robust visualizer tailored for transformer architectures. By processing training or test data, this project seeks to track and analyze changes in attention scores and other structural components of the transformer network, visualizing their evolution within individual input sentences during inference. The project involves developing innovative visualization conventions to compactly represent the high-dimensional neural network processes in a manner that is analytical and accessible, through the usage of static diagrams.