Recent News
Presented CaliDist Work at ICML 2026 Poster Session
Presented our latest paper "CaliDist: Calibrating Large Language Models via Behavioral Robustness to Distraction" at the ICML 2026 poster session in Seoul, South Korea!
Paper Accepted at ICML 2026!
Our paper "CaliDist: Calibrating Large Language Models via Behavioral Robustness to Distraction" has been accepted for publication!
Organized Bengali Code Generation Shared Task
Co-organized the shared task at BLP @ IJCNLP-AACL 2025 .
Research Focus
My research focuses on improving the reliability and reasoning capabilities of large language models, with an emphasis on uncertainty estimation, calibration, and reasoning reliability — through both post-hoc and post-training approaches. I am particularly interested in developing techniques that make model confidence meaningful and decision-relevant, especially under distribution shift and semantic perturbations. A central direction of my work investigates how model confidence and reasoning accuracy can be jointly improved — exploring both post-hoc adjustments and post-training interventions to bridge the gap between well-calibrated predictions and correct inference. More broadly, I aim to understand how model behavior changes under controlled perturbations and how these signals can be leveraged to build more robust, interpretable, and trustworthy AI systems.
Developing post-hoc and training-aware methods to improve the reliability of large language model confidence estimates, with a focus on calibration under distribution shift and perturbation-based evaluation. The goal is to ensure predictive probabilities are well-aligned with true correctness across diverse settings.
Exploring how to improve reasoning accuracy in large language models through both post-hoc and post-training methods. This includes perturbation-based inference techniques, lightweight verification signals, and training-time interventions to induce and select higher-quality reasoning chains.
Education
PhD in Computer Science
University of Illinois Chicago (UIC)
GPA: 4.00 / 4.00
Advisor: Professor Cornelia Caragea
BSc in Computer Science & Eng.
Islamic University of Technology (IUT)
GPA: 3.81 / 4.00
Honors & Certifications
Technical Toolkit
ML & Core Research
Languages
Data Processing
Databases
Web, Mobile & Tools
Publications
Selected publications and preprints.
CaliDist: Calibrating Large Language Models via Behavioral Robustness to Distraction
Mohammad Anas Jawad, Cornelia Caragea.
We introduce CaliDist, a novel framework that calibrates LLM confidence by measuring the model's behavioral robustness to semantic distractions.
Overview of BLP-2025 task 2: Code generation in bangla
Nishat Raihan, Mohammad Anas Jawad, Md Mezbaur Rahman, Noshin Ulfat, Pranav Gupta, Mehrab Mustafy Rahman, Santu Karmaker, Marcos Zampieri.
This paper details a shared task evaluating generative AI systems on producing executable Python code from Bangla prompts, highlighting that multi-method engineering pipelines consistently outperformed individual model solutions.
Renovo: Sensor-based visual assistive technology for physiotherapists in the rehabilitation of stroke patients with upper limb motor impairments
Mohammad Ridwan Kabir, Mohammad Ishrak Abedin, Mohaimin Ehsan, Mohammad Anas Jawad, Hasan Mahmud, Md Kamrul Hasan.
This study introduces a wearable IMU sensor system and a non-machine-learning mathematical framework that assists physiotherapists by providing real-time 3D tracking, visualization, and unbiased quantitative progress scores for stroke rehabilitation patients.
Professional Experience
Graduate Assistant (Research & Teaching)
University of Illinois Chicago (dl4nlpspace Lab)
- Conduct advanced core research targeting model calibration, uncertainty estimation, and logical reasoning pipelines in Large Language Models (LLMs).
- Instruct practical lab modules and host office hours for Discrete Math, Programming Practicum, and Systems Programming.
Lecturer
Islamic University of Technology (IUT)
- Designed academic curricula and delivered lectures for courses on Data Mining, Database Management Systems, Web Architecture, Visual Programming, and Object-Oriented Concepts.
- Actively collaborated on data mining research as a core investigator with the Network and Data Analysis Group (NDAG).
Software Development Intern
Nilavo Technologies Ltd.
Engineered and optimized backend relational database integrations with PostgreSQL under the Spring framework and refactored responsive dashboard views using AngularJS and Bootstrap.
Selected Projects
Context Optimization in Advanced RAG: Synthesize vs. Re-order
Developed a modular RAG pipeline comparing context synthesis and sentence-level re-ordering strategies using Llama-3.1-8B. Implemented DPR retrieval with FAISS and MS-MARCO re-ranking, improving Exact Match and semantic relevancy on HotpotQA.
Semi-Supervised Node Classification with LLMs
Designed a semi-supervised graph learning framework combining Llama-3.1-8B annotation with GCN-based pseudo-labeling. Implemented difficulty-aware filtering and confidence-based selection, outperforming supervised baselines on the Cora dataset under limited-label settings.
Wide & Deep Recommender System on XWine Dataset
Built a Wide & Deep recommendation model in TensorFlow for the XWines dataset using feature crosses and deep neural representations. Achieved improved validation MAE over standard baseline approaches.
Automated Essay Scoring & ML Classification
Developed an NLP-based essay evaluation system using linguistic and semantic features extracted with Word2Vec and Stanford Stanza. Trained MLP and Logistic Regression models for automated grading prediction.
Get in Touch
I'm always open to discussing research collaborations or opportunities.