Web IR NLP Group @ NUS
Web IR NLP Group @ NUS
Tour
News
People
Events
Gallery
Publications
Projects
Contact
Light
Dark
Automatic
Large Language Models
LLM-Augmented Health Community Notes
We improve health Community Notes through evidence-grounded augmentation, automated generation, and agents that learn from prior correction episodes.
Zihang Fu
,
Fanxiao Li
,
Jiaying Wu
,
Min-Yen Kan
CrowdNotes+ Paper
CrowdNotes+ Code & Data
EvoNote Paper
EvoNote Code & Data
Better with Experience: Self-Evolving LLM Agents for Evidence-Grounded Health Community Notes
EvoNote improves evidence-grounded health Community Notes by distilling prior correction episodes into reusable experience for claim analysis, evidence acquisition, and note writing.
Zihang Fu
,
Fanxiao Li
,
Jianyang Gu
,
Haonan Wang
,
Preslav Nakov
,
Bryan Hooi
,
Min-Yen Kan
,
Jiaying Wu
PDF
Cite
Code
Project
DOI
ACL Anthology
Cost-Efficient LLM-Driven Scientific Discovery
We study how to coordinate lightweight and frontier LLMs in evolutionary search for scientific discovery, staying competitive with frontier-model baselines at a fraction of the API cost.
Ha Minh Hieu
,
Tongyao Zhu
,
Xuan Long Do
,
Min-Yen Kan
SpecEvo Paper
SpecEvo Code
LLMs for Code: Faithful and Minimal Code Editing
We study whether LLM code repairs stay minimal and faithful to the original program, rather than only whether they pass the tests, and how models can be trained to edit code minimally.
Tongyao Zhu
,
Wei Hern Lim
,
Min-Yen Kan
Paper
Code & Data
SpecEvo: Speculative Evolution with Large Language Models for Cost-Efficient Scientific Discovery
A multi-LLM evolutionary search framework that pairs lightweight Speculators with a frontier Navigator, staying competitive with frontier-model baselines on mathematical discovery and systems optimization at up to 80% lower API cost.
Ha Minh Hieu
,
Tongyao Zhu
,
Xuan Long Do
,
Min-Yen Kan
PDF
Cite
Code
Project
DOI
arXiv
When Models Edit Too Much: On the Fidelity of Minimal Code Edits
LLM code repairs often pass the tests while rewriting far more code than needed. We measure this over-editing on a benchmark with known minimal patches and show that reinforcement learning can teach models to edit minimally.
Tongyao Zhu
,
Wei Hern Lim
,
Min-Yen Kan
PDF
Cite
Code
Project
DOI
ACL Anthology
arXiv
Why Is This Joke Offensive? Interpreting Offensiveness in Chinese Dark Humor Through Knowledge and Rhetoric
HellHumor-CN and HumorLens support the interpretation of offensive Chinese dark humor through background knowledge and rhetorical analysis.
Jiecheng Jiang
,
Herun Wan
,
Zihang Fu
,
Hai-Tao Yu
,
Min-Yen Kan
,
Jiaying Wu
PDF
Cite
Code
DOI
Context Window Scheduling for Efficient LLM Pretraining
Longer pretraining context windows are not always better. We study how to schedule the context window during LLM pretraining to balance long-context capability with training efficiency.
Tongyao Zhu
,
Qian Liu
,
Min-Yen Kan
Paper
arXiv
Code
Slides
Discursive Circuits: How Do Language Models Understand Discourse Relations?
We introduce a new task, Completion under Discourse Relation (CuDR), to reveal the underlying transformer circuits responsible for discourse understanding.
Yisong Miao
,
Min-Yen Kan
PDF
Cite
Code
Project
Slides
Reasoning and Planning for Large Language Models
ICLR 2025 Workshop on Reasoning and Planning for Large Language Models
28 Apr, 2025
Singapore Expo
Jiaying Wu
,
Min-Yen Kan
»
Cite
×