Hakeem Hannoon

Hi, I’m Hakeem, and I love Physics and AI. I want to be useful to other humans, understand more about the universe, and help answer some of the big questions.

AI Research · LLM Safety · Memory-Augmented LLMs January 2026 to April 2026

Research Fellow

Algoverse AI Research Program · Remote

Selected for a merit-based applied AI research cohort. First author of “Mitigating Over-Personalization in LLMs via Structured Memory,” accepted to multiple ICML workshops — developed the core methodology, built most of the codebase, and performed most of the computations, evaluations, and experimental results behind the paper.
- First author of “Mitigating Over-Personalization in LLMs via Structured Memory,” accepted to multiple ICML workshops.
- Developed the core methodology and built most of the codebase for structured-memory experiments across memory-augmented LLMs.
- Performed most computations, evaluations, results production, and improvements, including experiments showing dynamic memory partitioning reduced cross-domain leakage by 8.8 percent on average relative to the baseline.
LLM researchAI safetyMemory-augmented LLMsBenchmarkingPythonEvaluation pipelinesPrompt engineeringStructured memoryExperimental design

Read the paper
Scientific Computing · Quantum Chemistry · Big Data May 2025 to September 2025

Big Data Analyst

Advanced Physics Research Lab, University of Saskatchewan · Saskatoon, SK

Built a first-principles model predicting molecular toxicity from quantum-chemistry and DFT calculations, and the cloud pipelines that moved hundreds of gigabytes off the research supercomputing clusters.
- Built a first-principles model to predict toxicity of 25 PAH and PCB molecules using quantum chemistry and DFT calculations.
- Outperformed an existing model by over 30 percent when ranking chemical reactivity.
- Built cloud-based data pipelines for hundreds of gigabytes of scientific data from research supercomputing clusters.
Quantum chemistryDFTHigh-performance computingCloud pipelinesBig dataScientific MLPythonComputational physics
Mathematical Biology · HPC · Simulation May 2024 to October 2024

Research Assistant

Department of Mathematics & Statistics, University of Saskatchewan · Saskatoon, SK

Led a team of mathematicians, including two professors, studying the dynamic nature of biological systems and natural neural-network-inspired behavior using high-performance computing.
- Led a research project with mathematicians and two professors on dynamic biological systems and natural neural-network-inspired behavior.
- Developed five algorithms to capture collision behavior in diffusion-based biological systems with over 500,000 agents.
- Built models that identified critical behavior with 95 percent accuracy in higher dimensions, matching theoretical predictions.
Mathematical biologyDiffusion systemsHPCSimulationAlgorithm designScientific computingPythonApplied mathematics
Quantum Computing · Quantum Information June 2024 to August 2024

Research Assistant

quanTA, University of Saskatchewan · Saskatoon, SK

Worked with research students on foundational quantum information theory and near-term applications of quantum processing to materials discovery, health sciences, and energy.
- Worked on foundational quantum information theory and near-term quantum computing applications.
- Implemented Qiskit code for quantum algorithms and IBM Qiskit Summer School projects.
- Explored quantum processing applications in materials discovery, health sciences, and energy.
Quantum computingQiskitQuantum informationPythonIBM QiskitScientific computing

Publications

Papers & the thesis.

[ PAPER ] 2026 Accepted to multiple ICML workshops

Mitigating Over-Personalization in LLMs via Structured Memory

First author

This paper studies safety failures in memory-augmented conversational LLMs, especially cross-domain leakage and memory-induced sycophancy. It proposes inference-time structured memory representations that partition memories by domain instead of injecting all stored user memories as a flat list. Across seven models on PersistBench, the strongest dynamic partitioning method reduces cross-domain leakage by 8.8 percent on average relative to the baseline while preserving useful personalization.

LLM SafetyMemory-Augmented LLMsAI AlignmentPrivacyStructured MemoryPersistBench

Open Download
[ PAPER ] 2026 Undergraduate Honours Thesis

Honours Thesis

Author

Research on symmetry methods for solving non-linear partial differential equations — using the structure of a system's symmetries to construct exact and reduced solutions where direct integration fails.

Mathematical PhysicsPartial Differential EquationsSymmetry MethodsApplied MathematicsScientific Computing

Open Download
[ PAPER ] 2026 Accepted to multiple ICML workshops

Mitigating Over-Personalization in LLMs via Structured Memory

First author

This paper studies safety failures in memory-augmented conversational LLMs, especially cross-domain leakage and memory-induced sycophancy. It proposes inference-time structured memory representations that partition memories by domain instead of injecting all stored user memories as a flat list. Across seven models on PersistBench, the strongest dynamic partitioning method reduces cross-domain leakage by 8.8 percent on average relative to the baseline while preserving useful personalization.

LLM SafetyMemory-Augmented LLMsAI AlignmentPrivacyStructured MemoryPersistBench

Open Download
[ PAPER ] 2026 Undergraduate Honours Thesis

Honours Thesis

Author

Research on symmetry methods for solving non-linear partial differential equations — using the structure of a system's symmetries to construct exact and reduced solutions where direct integration fails.

Mathematical PhysicsPartial Differential EquationsSymmetry MethodsApplied MathematicsScientific Computing

Open Download
[ PAPER ] 2026 Accepted to multiple ICML workshops

Mitigating Over-Personalization in LLMs via Structured Memory

First author

This paper studies safety failures in memory-augmented conversational LLMs, especially cross-domain leakage and memory-induced sycophancy. It proposes inference-time structured memory representations that partition memories by domain instead of injecting all stored user memories as a flat list. Across seven models on PersistBench, the strongest dynamic partitioning method reduces cross-domain leakage by 8.8 percent on average relative to the baseline while preserving useful personalization.

LLM SafetyMemory-Augmented LLMsAI AlignmentPrivacyStructured MemoryPersistBench

Open Download
[ PAPER ] 2026 Undergraduate Honours Thesis

Honours Thesis

Author

Research on symmetry methods for solving non-linear partial differential equations — using the structure of a system's symmetries to construct exact and reduced solutions where direct integration fails.

Mathematical PhysicsPartial Differential EquationsSymmetry MethodsApplied MathematicsScientific Computing

Open Download
[ PAPER ] 2026 Accepted to multiple ICML workshops

Mitigating Over-Personalization in LLMs via Structured Memory

First author

This paper studies safety failures in memory-augmented conversational LLMs, especially cross-domain leakage and memory-induced sycophancy. It proposes inference-time structured memory representations that partition memories by domain instead of injecting all stored user memories as a flat list. Across seven models on PersistBench, the strongest dynamic partitioning method reduces cross-domain leakage by 8.8 percent on average relative to the baseline while preserving useful personalization.

LLM SafetyMemory-Augmented LLMsAI AlignmentPrivacyStructured MemoryPersistBench

Open Download
[ PAPER ] 2026 Undergraduate Honours Thesis

Honours Thesis

Author

Research on symmetry methods for solving non-linear partial differential equations — using the structure of a system's symmetries to construct exact and reduced solutions where direct integration fails.

Mathematical PhysicsPartial Differential EquationsSymmetry MethodsApplied MathematicsScientific Computing

Open Download
[ PAPER ] 2026 Accepted to multiple ICML workshops

Mitigating Over-Personalization in LLMs via Structured Memory

First author

This paper studies safety failures in memory-augmented conversational LLMs, especially cross-domain leakage and memory-induced sycophancy. It proposes inference-time structured memory representations that partition memories by domain instead of injecting all stored user memories as a flat list. Across seven models on PersistBench, the strongest dynamic partitioning method reduces cross-domain leakage by 8.8 percent on average relative to the baseline while preserving useful personalization.

LLM SafetyMemory-Augmented LLMsAI AlignmentPrivacyStructured MemoryPersistBench

Open Download
[ PAPER ] 2026 Undergraduate Honours Thesis

Honours Thesis

Author

Research on symmetry methods for solving non-linear partial differential equations — using the structure of a system's symmetries to construct exact and reduced solutions where direct integration fails.

Mathematical PhysicsPartial Differential EquationsSymmetry MethodsApplied MathematicsScientific Computing

Open Download
[ PAPER ] 2026 Accepted to multiple ICML workshops

Mitigating Over-Personalization in LLMs via Structured Memory

First author

This paper studies safety failures in memory-augmented conversational LLMs, especially cross-domain leakage and memory-induced sycophancy. It proposes inference-time structured memory representations that partition memories by domain instead of injecting all stored user memories as a flat list. Across seven models on PersistBench, the strongest dynamic partitioning method reduces cross-domain leakage by 8.8 percent on average relative to the baseline while preserving useful personalization.

LLM SafetyMemory-Augmented LLMsAI AlignmentPrivacyStructured MemoryPersistBench

Open Download
[ PAPER ] 2026 Undergraduate Honours Thesis

Honours Thesis

Author

Research on symmetry methods for solving non-linear partial differential equations — using the structure of a system's symmetries to construct exact and reduced solutions where direct integration fails.

Mathematical PhysicsPartial Differential EquationsSymmetry MethodsApplied MathematicsScientific Computing

Open Download

Websites

Things I've built.

Education

Where the training happened.

Awards, activities, coursework, and conferences.

University of Saskatchewan

Honours, Mathematical Physics & Computer Science

Sep 2023 – 2026 High Honours

Mathematical physicsQuantum mechanicsApplied mathematicsComputer scienceArtificial intelligenceScientific computing

Awards & Scholarships

University of Saskatchewan Awards for Undergraduate Students
2026
Undergraduate Arts & Science Scholarship
2026
International Students at Risk Award
2025
Student Research Award
Spr/Sum 2025
2× Dean's List
2024 & 2025
2× Honours Scholarship
2024 & 2025
International Undergraduate Student Award
2025
SIOS Students Scholarship
2025
International Students at Risk Award
2024
Arts & Science Student Union Scholarship
2024
quanTA Student Research Award
Spr/Sum 2024
Undergraduate Student Research Award
Spr/Sum 2024

Activities

Rugby Player

University of Saskatchewan Rugby Team
Member

University of Saskatchewan Rocketry Team
Member

Arts & Science Student Union
Volunteer Tutor & Coach

Community & University Gym
- Taught Quantum Mechanics, Data Structures, Vector Calculus, Calculus I & II, Physics I–III, and C programming.
- Coached weightlifting and recovery technique at the university gym.

Selected coursework

Linear Algebra IIProbability TheoryAdvanced Quantum MechanicsPartial Differential EquationsIntroduction to AIReal Analysis IComplex AnalysisIntroduction to TopologyIntro Nuclear & Particle PhysicsQuantum MechanicsApplied Differential EquationsQuantum ComputingAdvanced ElectromagnetismDeep Learning & ApplicationsAutomata & Formal LanguagesElectromagnetic Fields & Circuits

Conferences & summer schools

Physics Undergraduate Students Conference 2024 — UBC, Oct 2024
MMI Summit for Bio Mathematics — UBC, Sep 2024
Qiskit Global Summer School 2024 — IBM, Jun 2024
PIMS Quantum Computing Summer School — USask, May 2024

Also: Google Advanced Data Analytics Certificate (2025) · AI with Python for Beginners (2025)

University of Waterloo

Master of Data Science & Artificial Intelligence

Incoming · 2026

Data scienceArtificial intelligenceMachine learning

Contact

Let’s talk physics, AI, or what comes next.

You can reach out about research, collaboration, or any question you have in mind.

Hakeem Hannoon

Research Fellow

Big Data Analyst

Research Assistant

Research Assistant

Papers & the thesis.

Mitigating Over-Personalization in LLMs via Structured Memory

Honours Thesis

Mitigating Over-Personalization in LLMs via Structured Memory

Honours Thesis

Mitigating Over-Personalization in LLMs via Structured Memory

Honours Thesis

Mitigating Over-Personalization in LLMs via Structured Memory

Honours Thesis

Mitigating Over-Personalization in LLMs via Structured Memory

Honours Thesis

Mitigating Over-Personalization in LLMs via Structured Memory

Honours Thesis

Things I've built.

Aether Field

Spotter

Aether Field

Spotter

Aether Field

Spotter

Aether Field

Spotter

Aether Field

Spotter

Aether Field

Spotter

Where the training happened.

University of Saskatchewan

Awards & Scholarships

Activities

Selected coursework

Conferences & summer schools

University of Waterloo

Let’s talk physics, AI, or what comes next.