Nusrat Lia

Latest News

Jun 27, 2026

Bangladesh won Three GOLD in the Asia Pacific Olympiad in AI other Top contenders being CHINA, JAPAN, RUSSIA, AUSTRALIA...(Proud Instructor Moment!) see

Jun 18, 2026

Preprint available for LRE which beats Microsoft's latest work ACON reducing peak context size by ~52%, with zero added neural cost and matches the accuracy of keeping the entire history

May 26, 2026

AgentCollabBench accepted at FAGEN@ICML 2026 (Non archival) - 900 tasks that catch when a multi-agent LLM team's final answer is right but collaborative reasoning was broken.

May 9, 2026

Served as a Judge for International AI Olympiad (BD)

May 9, 2026

Served as a Judge at a National Hackathon 2026 (NDCIT)

Recent Work

Fatima Institute

Research Fellow

Fatima Institute for Global AI Research, California

Apr 2026 - Present

Aramco-Ithra

Research Intern

Aramco-Ithra

Mar 2025 - Mar 2026

ICITAP

WPA Software Engineer Intern

US Department of Justice - ICITAP

Jul 2024 - Dec 2024

IAIO

Instructor and Judge

International AI Olympiad

Sep 2025 - Present

Education

University of Dhaka

University of Dhaka

B.Sc. in Software Engineering

May 2026 · CGPA 3.8/4 (Top 5)

Courses & Teaching

Building Small Language Model: From Foundations to Bangla Financial Text Generation

Building Small Language Model

University of Dhaka

I teach this course offline at the Institute of Information Technology, University of Dhaka. To enroll in the most updated course classes, contact BARTA Lab (barta-research-lab.github.io).

Collaborations

University of Utah

University of Utah

UMBC

UMBC

University of Virginia

University of Virginia

World Health Organization

WHO

Stony Brook University

Stony Brook University

Seattle Children's Research Institute

Seattle Children's

University of Washington

Univ. of Washington

SUNY Old Westbury

SUNY Old Westbury

Toronto Metropolitan University

Toronto Metro Univ.

University of Geneva

University of Geneva

Imperial College London

Imperial College London

Purdue University

Purdue University

Featured Projects

Privacy Preserved Federated Code Recommendation System
Privacy Preserved Federated Code Recommendation System
AI/ML
This is a distributed RAG (Retrieval-Augmented Generation) system that enables organizations to share code recommendations while preserving privacy

Latest Blogs

When Good Agents Make Bad Collaborators

Hidden Behavioral and Reliability Failures in Multi-Agent LLMs Beyond Task Accuracy

May 12Read

Hi, I'm Lia

My work lies in natural language processing, human-centered applications and secured decentralized systems, with experience in large-scale software and LLM development

I was a Research Intern at Aramco-Ithra, collaborating with global institutions including WHO, Stony Brook Medicine, University of Washington, University of Geneva, University of Tokyo and research institutes from 35 countries. Previously, I worked with the United States Department of Justice - ICITAP, designed a platform for secure crowdsourced crime reporting in low-connectivity areas, leveraging custom NLP pipelines, geospatial and predictive models.

I recently graduated (Top 5) with a degree in Software Engineering (concentrating in AI and NLP) from the University of Dhaka where I work in BARTA Lab. There, I worked on low-resource and small-language-model development. I serve as an instructor at BARTA, where I teach a language model building course . I am also serving as an Instructor for International AI Olympiad, teaching AI Recommender Systems, and a Judge at BDAIO.

Entrepreneurially, I was a founding member of Perspectivity - Drishtikon, the first real-time AI news aggregator for Bangla, featuring multi-axis bias detection that empower citizens with research agents.

And... I paint.

Research

AgentCheck: A Reproduce–Intervene–Mitigate Workbench for LLM Agents over MCP

AgentCheck: A Reproduce–Intervene–Mitigate Workbench for LLM Agents over MCP

Aritra Mazumder; Nusrat Jahan Lia
2026
Learning What Not to Forget: Long-Horizon Agent Memory from a Few Kilobytes of Learning

Learning What Not to Forget: Long-Horizon Agent Memory from a Few Kilobytes of Learning

Nusrat Jahan Lia; Aritra Mazumder
2026
AGENTCOLLABBENCH: Diagnosing When Good Agents Make Bad Collaborators

AGENTCOLLABBENCH: Diagnosing When Good Agents Make Bad Collaborators

Aritra Mazumder; Shubhashis Roy Dipta; Nusrat Jahan Lia; et al.
2026
Register Shifts Break LLM Safety: A Bengali Benchmark with Culturally Grounded Harms

Register Shifts Break LLM Safety: A Bengali Benchmark with Culturally Grounded Harms

Naymul Islam, Nusrat Jahan Lia, Shubhashis Roy Dipta (equal first authors) et al.
2026
Cross-Lingual Sentiment Misalignment: Auditing Multilingual Language Models for Inversion Risk, Dialectal Representation, and Affective Stability

Cross-Lingual Sentiment Misalignment: Auditing Multilingual Language Models for Inversion Risk, Dialectal Representation, and Affective Stability

Nusrat Jahan Lia; Shubhashis Roy Dipta
mellm @ ACL 2026; Published: ACL Anthology2026
Read Between the Lines: A Benchmark for Uncovering Political Bias in Bangla News Articles

Read Between the Lines: A Benchmark for Uncovering Political Bias in Bangla News Articles

Nusrat Jahan Lia; Shubhashis Roy Dipta, PhD; Dr. Abdullah Khan Zehady; Naymul Islam; Madhusodan Chakraborty; Abdullah Al Wasif
Accepted: AACL IJCNLP BLP; Published: ACL Anthology2025
Exploring Cross-Lingual Knowledge Transfer via Transliteration-Based MLM Fine-Tuning for Critically Low-resource Chakma Language

Exploring Cross-Lingual Knowledge Transfer via Transliteration-Based MLM Fine-Tuning for Critically Low-resource Chakma Language

Adity Khisa; Nusrat Jahan Lia; Tasnim Mahfuz Nafis; Zarif Masud; Tanzir Pial, PhD; Dr.Shebuti Rayana; Dr.Ahmedul Kabir
Accepted: AACL IJCNLP BLP; Published: ACL Anthology2025

Statistical Inference & Data Provenance Across 35 Countries

These were done during early undergrad. I was responsible for discovering patterns and findings across 35 countries. Some of which made it to publication, some others made it to project-level implementation, or presentations in places like WHO headquarters.

Adult Attitudes about School Smartphone Bans: A Global Survey of 35 Countries

Adult Attitudes about School Smartphone Bans: A Global Survey of 35 Countries

Dimitri A. Christakis, MD, MPH; Nusrat Jahan Lia; Lauren Hale, PhD
Accepted, Published: The Journal of American Medical Association; doi: 10.1001/jamapediatrics.2025.57362025
Does Gaming Disorder Symptom Status Predict Poorer Sleep Quality?

Does Gaming Disorder Symptom Status Predict Poorer Sleep Quality?

Nusrat Jahan Lia; Lauren Hale, PhD; Justin Thomas, PhD; Dimitri A. Christakis, MD, MPH; Mamunar Rashid, PhD
Accepted: World Sleep 2025, Singapore2025
Does Spending "Too Much Time Online" Predict Sleep Health and Mental Health?

Does Spending "Too Much Time Online" Predict Sleep Health and Mental Health?

Lauren Hale, PhD; Nusrat Jahan Lia; Sohailul Islam Alvi; Gina Marie Mathew, PhD; Dimitri A. Christakis, MD, MPH; Mamunar Rashid, PhD; Yasmin Aljedawi; Melisa Valle, PhD
Accepted: Association of Professional Sleep Societies. Seattle, Washington, USA2025
International Public Opinion on Digital Media Use for Youth and Schools

International Public Opinion on Digital Media Use for Youth and Schools

Lauren Hale, PhD; Nusrat Jahan Lia; Sohailul Islam Alvi; Gina Marie Mathew, PhD; Dimitri A. Christakis, MD, MPH; Mamunar Rashid, PhD; Yasmin Aljedawi; Melisa Valle, PhD
Accepted: Digital Media and Developing Minds International Scientific Congress, Washington DC2025

Ongoing Work

A Difficulty-Parametric Diagnostic Benchmark for Tool-Using Agents with 11 Failure Attribution, including security collapse in a Simulated World

UMBC

When Confidence Becomes Misleading: Runtime Monitoring in Multi-Step LLM Agents

University of Utah
WOOW

Director

WOOW (Work for Orientation & Organizing the World)

2022 - Present

We work to support underprivileged families, orphanages, and children with special needs through resources, medical assistance and community development programs. A preview is here: