Abdullah Hashmat

Abdullah Hashmat

Research Fellow

Max Planck Institute for Software Systems (MPI-SWS)

Kaiserslautern, Germany

B.S. Computer Science, LUMS (2025)

abdullah.hashmat02@gmail.com

About

I am an AI researcher and Computer Science graduate focused on the intersection of Machine Learning, Scientific AI, and Computational Biology. My research experience includes LLMs, speech technologies for low-resource languages, benchmark development, molecular property prediction, and generative models for drug discovery.

Currently, I am a Research Fellow at the Max Planck Institute for Software Systems (MPI-SWS) in Kaiserslautern, Germany, exploring Agentic Systems and Scientific AI. Previously, at DFKI, I worked on bioinformatics and state-space modeling techniques for molecular representation learning.

I am driven by the goal of leveraging AI to advance science, healthcare, and multilingual technologies. For my final year thesis, I created PakBBQ, a bias benchmarking QA dataset with 17,180 Urdu/English examples across 8 sociocultural categories. I evaluated 6 multilingual LLMs, revealing cross-linguistic bias gaps and effective bias mitigation strategies. This work culminated in a paper accepted at EMNLP 2025 main conference (Core: A*).

News & Updates

June 2026

Joined the Max Planck Institute for Software Systems (MPI-SWS) as a Research Fellow

November 2025

Paper accepted at EMNLP 2025 Main Conference (Core: A*)

June 2025

Started as Research Assistant at DFKI Germany

May 2025

Graduated from LUMS with B.S. in Computer Science

January 2025

Appointed as Teaching Assistant for CS5302 Foundations of Generative AI

January 2025

Appointed as Teaching Assistant for AI600 Machine Learning (MS in AI)

September 2024

Appointed as Teaching Assistant for CS535 Machine Learning

September 2023

Joined AI in Healthcare Initiative (AIHI) as a Research Assistant

September 2023

Joined Center for Speech and Language Technologies (CSaLT) as a Research Assistant

September 2023

Appointed as Teaching Assistant for CS100 Computational Problem Solving

Publications

PakBBQ: A Culturally Adapted Bias Benchmark for QA

EMNLP 2025 (Main Conference, Core: A*)

With the widespread adoption of Large Language Models (LLMs), it's crucial to ensure fairness across all user communities. Most LLMs are trained on Western-centric data, neglecting low-resource languages and regional contexts. To address this, we introduce PakBBQ — an extension of the Bias Benchmark for Question Answering (BBQ) dataset — featuring 214 templates and 17,180 QA pairs in English and Urdu across 8 bias dimensions: age, disability, appearance, gender, socio-economic status, religion, regional affiliation, and language formality.

Our experiments show: (i) a 12% accuracy gain with disambiguation, (ii) stronger counter-bias behaviors in Urdu, and (iii) framing effects reducing stereotypes in negatively posed questions.

Research Interests

Large Language Models

  • Benchmark Development
  • LLM Biases and Fairness
  • Reasoning & Robustness

Scientific AI

  • Agentic Systems
  • Generative Models for Drug Discovery
  • AI for Scientific Discovery

Computational Biology

  • Molecular Property Prediction
  • Molecular Representation Learning
  • State-Space Modeling

Low-Resource Speech & NLP

  • Speech Technologies
  • Multilingual Evaluation
  • Fair & Ethical AI

Selected Projects

PakBBQ
2025 FYP

PakBBQ: Bias Benchmark

Culturally adapted QA benchmark for multilingual LLM bias evaluation. 17,180 examples in Urdu/English across 8 bias categories.

View Project →
Knowledge Distillation
2025 Research

Bias Mitigation KD

Confidence-aware knowledge distillation achieving balanced shape/texture bias. Improved OOD generalization on vision models.

View Project →
AI for All
2024 Low-Resource NLP

AI for All

Evaluating LLMs in Sindhi, Pashto, and Urdu. Bias and jailbreak testing on lightweight transformers.

View Project →
AI Stock Insights
2024 AWS & LLMs

AI Stock Insights

AWS-based platform for stock analysis. S3, Lambda, ECS, Postgres with LLM-powered chatbot.

View Project →
Data Science Chatbot
2024 Agentic AI

Data Science Chatbot

End-to-end AI chatbot for EDA and data cleaning. Fine-tuned GPT-3.5, deployed on Hugging Face.

View Project →
Unity Game
2024 Game Dev

Platformer Game

2D platformer in Unity with 10+ levels. Hand-drawn pixel art, dynamic lighting, puzzle mechanics.

View Project →

Skills

Programming

  • Python
  • JavaScript / TypeScript
  • C++ / C / C#
  • Matlab
  • SQL

AI & ML

  • PyTorch
  • TensorFlow
  • scikit-learn
  • Transformers
  • NLTK

Data & Tools

  • NumPy & Pandas
  • GeoPandas
  • DynamoDB
  • Matplotlib & Seaborn

Development

  • React & MERN Stack
  • Redux
  • Tailwind CSS

Cloud & DevOps

  • AWS
  • Google APIs
  • Neon Postgre

Other

  • Gradio
  • LaTeX
  • Unity
  • Hugging Face