Ahmed Karim

PhD student

UCL Robot Perception and Learning Lab
EPSRC CDT in Cyber-Physical Risk

Ahmed Karim

About

I'm a PhD student at UCL's Robot Perception and Learning Lab, supervised by Prof. Dimitrios Kanoulas. My research focuses on foundational world models for robotics: how autonomous systems learn to perceive, predict, and act in environments they have never seen before.

I work at the intersection of world models, open-ended learning, and embodied AI, with applications in autonomous systems for extreme environments such as disaster response, planetary exploration, and deep sea. Alongside this, I do research on language models, focusing on reliability, hallucination, and the information-theoretic structure behind their failures.

I'm also a Research Fellow at Hassana Labs. Previously, I was a Machine Learning Engineer at the London Stock Exchange Group (Advanced Data Analytics).

News

  • Oct 2026Started my PhD at the UCL Robot Perception and Learning Lab.
  • Sep 2026New preprint, under review at ICLR 2027: Robot World Models Are Not Invariant to How the Actions Are Written.
  • Apr 2026Paper accepted at ICML 2026: Predictable Compression Failures: Why Language Models Actually Hallucinate.
  • 2026Awarded EPSRC TechExpert PhD studentship at the UCL CDT in Cyber-Physical Risk.
  • 2025Paper accepted at EMNLP 2025: Beyond the Score.

Publications

A robot given the same directions written two ways, one leading to Earth and the other to Mars

Robot World Models Are Not Invariant to How the Actions Are Written

A. Karim, L. Chlon

Under review at ICLR 2027paper

Two probability distributions illustration

Predictable Compression Failures: Why Language Models Actually Hallucinate

L. Chlon, A. Karim, M. Chlon

ICML 2026papercode

Conformal prediction set illustration

Beyond the Score: Uncertainty-Calibrated LLMs for Automated Essay Assessment

A. Karim, Q. Wang, Z. Yuan

EMNLP 2025paper

Multi-layer attention trace illustration

Attention Deficits in Language Models: Causal Explanations for Procedural Hallucinations

A. Karim, F. Sheaib, Z. Khamis, M. Chlon, J. Awada, L. Chlon

arXiv preprintpaper

Open Source

  • Berry: OpenAI-compatible Hallucination Risk Calculator for LLM reliability. 1,600+ stars.
  • Mezzanine: toolkit for diagnosing reliability and structure failures in modern ML systems. 150+ stars.

Service

  • Program Committee, Multi-LLL 2026 (Workshop on Multi-modality and Multi-lingual Language Learning) @ AACL-IJCNLP 2026.

Résumé

  1. Oct 2026 – present

    UCL · PhD in Computer Science

    Foundational world models for robotics, supervised by Prof. Dimitrios Kanoulas.

  2. 2025 – present

    Hassana Labs · Research Fellow (part-time)

    LLM reliability and hallucination research. Co-author on ICML 2026 paper. Co-built Berry, an open-source hallucination risk calculator (1.6k+ ★).

  3. 2024 – 2026

    London Stock Exchange Group · Machine Learning Engineer

    Advanced Data Analytics team. End-to-end ML systems: time-series forecasting (ARIMAP, 20% RMSE reduction), RAG-based internal chatbot, NLP+LLM incident-triage pipeline on AWS SageMaker.

  4. 2024

    King's College London · Researcher

    Uncertainty-calibrated LLMs for automated essay scoring; led to EMNLP 2025 paper.

  5. 2023

    London Stock Exchange Group · DevOps Engineer Intern

    Cloud migration with Helm, Terraform, Kubernetes. AWS training in LLMs.

  6. 2022

    Imperial College London · Undergraduate Researcher (UROP)

    Machine learning for wildfire prediction and propagation at the Leverhulme Centre for Wildfires. £2,500 funded.

  7. 2020 – 2023

    Queen Mary, University of London · BSc Computer Science (First Class)

    Final-year project: model-based reinforcement learning with curiosity-driven exploration.