Fardin Ahsan Sakib

PhD Candidate, Computer Science
George Mason University

Graduate Research Assistant @ George Mason University

I am a doctoral researcher in the Department of Computer Science at George Mason University and a member of the MasonNLP Research Group. My research spans the large language model lifecycle — interpretability, post-training, and evaluation — aimed at understanding how these models behave and making them dependable for real-world, high-stakes use such as clinical NLP and healthcare, with work published at venues like ACL and EMNLP. A central thread is AI safety: making models and the systems built on them fairer, more robust, and better aligned — from mitigating bias and reliability failures in LLMs to AI safety in multi-agent systems. Through research internships at Comcast, Fujitsu Research, and Amazon, I extend this to applied agentic AI — including agentic memory, agent evaluation, coding agents, and tool-calling workflows.

AI Safety Multi-Agent Systems Agentic Memory Agent Evaluation Large Language Models Post-Training Mechanistic Interpretability Clinical NLP
LinkedIn Google Scholar
Fardin Ahsan Sakib

Recent News

Experience

AI Research Intern

Comcast (Xfinity Assistant, Applied AI)
June 2026 – August 2026 Washington, DC
  • Designed and built cross-session memory for a multi-agent LLM assistant that adapts agent behavior for repeat customers, raising prior-context acknowledgment from 22% to 89%.
  • Created an LLM-as-judge evaluation pipeline to measure memory quality and agent behavior, and used it to improve fact coverage from 45% to 80%.

Agentic AI Research Intern

Fujitsu Research of America
November 2025 – January 2026 Santa Clara, CA
  • Engineered the candidate-selection stage for test-time scaling of LLM coding agents on SWE-Bench Verified, picking the best of 8 agent-generated patches using cross-agent test results.
  • Trained an LLM verifier (87% accuracy) and compared it with test-based selection, showing that executable test evidence ranks candidate patches more reliably.

Applied Research Scientist Intern

Amazon (AWS Support)
June 2025 – August 2025 Seattle, WA
  • Built a multi-step tool-calling benchmark from real AWS support queries that require chained, cross-service API calls, with a simulation server that replays recorded API responses for reproducible, cost-free agent evaluation.
  • Benchmarked frontier LLM agents using step-level metrics and a validated LLM-as-judge, showing the best model resolved only ~20% of real support queries.

Applied Research Intern

Brillient Corporation
May 2024 – August 2024 Reston, VA
  • Implemented query distillation, chunking, and re-ranking for a retrieval-augmented generation (RAG) system, improving retrieval relevance by 22% (MAP) and cutting processing time by 35%.
  • Fine-tuned domain-specific LLMs and deployed the system across medical, HR, and IRS domains, reaching 85% user satisfaction.

Graduate Research Assistant

George Mason University
August 2021 – Present Fairfax, VA
  • Designed evaluation frameworks exposing reliability failures in LLMs, showing models rely on spurious features such as demographic cues, with up to 37-point error gaps across groups.
  • Developed an inference-time correction method that reads the model's layer-by-layer prediction trajectory to decide which predictions to fix, correcting up to 79% of errors with low break rates across three model families and four datasets, with no retraining.
  • Proposed a training-free defense for multi-agent LLM systems against prompt injection, reducing attack success by over 80% with under one point of accuracy loss and outperforming seven existing defenses.
  • Led labs and discussions for 100+ students, providing individualized feedback on assignments and exams.

Machine Learning and AI Intern

Brillient Corporation
May 2023 – August 2023 Reston, VA
  • Engineered and fine-tuned NLP models for QA, summarization, emotion detection, and readability analysis, improving performance by 12–18% (F1, ROUGE, accuracy) for IRS and USCIS projects.
  • Built an end-to-end ML pipeline from model development to AWS deployment and API creation (SageMaker, Lambda, API Gateway), reducing deployment time by 40%.

Selected Publications

Trajectory-Guided Steering: Mitigating Spurious Feature Bias in Clinical Classification

Sakib, Fardin Ahsan; Ziwei Zhu; Özlem Uzuner
Under review

The Model Knows Better: Defending Medical Multi-Agent Systems Against Prompt Injection

Sakib, Fardin Ahsan; Ziwei Zhu; Özlem Uzuner
Under review

Spurious Correlations and Beyond: Understanding and Mitigating Shortcut Learning in SDOH Extraction with Large Language Models

Sakib, Fardin Ahsan; Ziwei Zhu; Karen Trister Grace; Meliha Yetişgen; Özlem Uzuner
ACL 2025
Paper

To token or not to token: A Comparative Study of Text Representations for Cross-Lingual Transfer

Rahman, Md Mushfiqur; Fardin Ahsan Sakib; Fahim Faisal; Antonios Anastasopoulos
MRL @ EMNLP 2023
Paper

Intent Detection and Slot Filling for Home Assistants: Dataset and Analysis for Bangla and Sylheti

Sakib, Fardin Ahsan; A. H. M. Rezaul Karim; Saadat Hasan Khan; Md Mushfiqur Rahman
BLP-2023 @ EMNLP 2023
Paper

MASON-NLP at eRisk 2023: Deep Learning-Based Detection of Depression Symptoms from Social Media Texts

Sakib, Fardin Ahsan; Ahnaf Atef Choudhury; Özlem Uzuner
CLEF 2023
Paper
View all publications on Google Scholar →

Awards & Honors

Distinguished Academic Achievement Award

George Mason University
2024

NSF Research Trainee (NRT) Fellowship

Center for Adaptive Systems of Brain-Body Interactions
2023

Academic Service

Reviewer — ACL Rolling Review (ARR)

Association for Computational Linguistics

Reviewer — Deep Learning for Code (DL4C)

Workshop @ NeurIPS

Reviewer — Bangla Language Processing (BLP)

Workshop co-located with EMNLP / IJCNLP-AACL

Technical Expertise

Research Areas

Large Language Models Foundation Models Generative AI Mechanistic Interpretability Inference-time Steering Post-training Bias Mitigation AI Safety Model Alignment Multi-Agent Systems Prompt Injection Adversarial Robustness Clinical NLP Healthcare AI

Methods

Fine-tuning LoRA RLHF DPO RAG Verifier Models Reward Modeling Model Evaluation Benchmarking LLM-as-Judge Agentic AI Coding Agents Tool Use Test-time Scaling Prompt Engineering

Frameworks & Tools

PyTorch Hugging Face Transformers LangChain Multi-GPU Training HPC Clusters Vector Databases Model Context Protocol (MCP)

Programming

Python C++ Java SQL

Cloud & MLOps

AWS (SageMaker, Lambda, API Gateway) Azure Docker Kubernetes Git

Education

Ph.D. in Computer Science

George Mason University
August 2021 – Present Fairfax, VA · GPA 3.95 / 4.00

M.S. in Computer Science

George Mason University
August 2021 – May 2024 Fairfax, VA

B.Sc. in Computer Science and Engineering

Islamic University of Technology
January 2017 – March 2021 Dhaka, Bangladesh · GPA 3.74 / 4.00

Press

George Mason University 08/2025 Article

AI's Blind Spots: One PhD Student's Clear Vision