Research Stream

Filter
Feb 2025 – Nov 2026 Preprint
Preprint on arXiv IEEE TNNLS

Sentence-Level Attribution for Black-Box Large Language Model Outputs via Energy-Based Surrogates

Maryam Rezaee (1st Author), Pooriya Safaei (2nd Author), Maryam Asgarinezhad (3rd Author), Fatemeh Seyyedsalehi (Corresponding Author) •TGML Lab, Sharif University of Technology

A model-agnostic, post-hoc attribution framework (ESCI) for black-box Large Language Models accessed strictly via APIs. ESCI uses an Energy-Based Model surrogate to map prompt-response consistency, then trains a lightweight interpreter network to deliver sentence-level importance scores in constant time without inference-time API queries.

Black-Box LLMs Post-Hoc Interpretability Energy-Based Models Sentence Attribution XAI
Sep 2024 – Aug 2026 Print
Sharif M.Sc. Thesis

Interpretability in Generative Models: Investigating the Mechanisms Behind Output Generation in Large Language Models

Maryam Rezaee (Master's Researcher) •TGML Lab, Sharif University of Technology

A master's thesis investigating interpretability mechanisms in generative language models. Features a comprehensive literature review categorizing intrinsic, post-hoc, concept-based, and mechanistic methods alongside Energy-Based Models in NLP, and develops an amortized post-hoc framework for black-box input-output attribution.

Generative Models Model Interpretability Black-Box Auditing Energy-Based Models Mechanistic Interpretability
May 2025 – Present Active
Undergoing Final Edits

Arc-CBM: Prototype-Grounded Concept Bottlenecks with Manifold-Grounded Visual Auditing for Medical Image Analysis

Mostafa Abdolmaleki (1st Author), Maryam Rezaee (2nd Author), Fatemeh Seyyedsalehi (Corresponding Author) •TGML Lab, Sharif University of Technology

A prototype-grounded concept bottleneck model for chest radiography that defines clinical concepts using L2-projected positive and negative prototypes from report-labeled support images, paired with ArcMAP for manifold-projected visual auditing without extra encoder passes.

Medical Imaging Concept Bottleneck Models XAI Prototype Grounding Manifold Learning
May 2025 – Present Active
Improving Method

Concept-Based Interpretability for Retrieval-Augmented Generation (RAG) Systems on Large Language Models

MohammadHossein MahdiKhani (Research Mentee), Maryam Rezaee (Research Mentor), Fatemeh Seyyedsalehi (Lab Director) •TGML Lab, Sharif University of Technology

Mentoring undergraduate research on concept-based interpretability for Retrieval-Augmented Generation (RAG) systems, designing methodologies with Concept Activation Vectors and Automatic Concept-based Explanations to verify factual inheritance and output transparency.

LLMs RAG XAI Concept Activation Vectors Interpretability
Jan 2021 & JUl 2025 Preprint
B.Sc. Project

A Neurosymbolic Architecture for Coherent Narrative Generation

Maryam Rezaee (Undergraduate Researcher) •University of Isfahan

A neurosymbolic director-actor architecture designed to resolve the horizon problem in large language models by formalizing canonical narrative arcs through signal processing envelopes, gated finite state machines, and vector centroid archetypes.

Neurosymbolic AI Narrative Generation Signal Processing Finite State Machines Active Inference
Feb 2023 – Sep 2023 Other
B.Sc. Project

Understanding the Application of Machine Learning in Cell Signalling Pathway Analysis and Developing a DEG Analysis Handbook

Maryam Rezaee (Undergraduate Researcher) •University of Isfahan

An investigation into machine learning techniques for biological network analysis and gene expression profiling, accompanied by a comprehensive technical handbook and implementation roadmap for Differentially Expressed Gene (DEG) analysis.

Bioinformatics Machine Learning Differential Gene Expression Cell Signalling Computational Biology
Aug 2016 – Feb 2017 Other
KaUMS Research Award

Foundational Research in Cognitive Neuroscience

Maryam Rezaee, Kowsar Kheirkhah, Fatemeh Jafari, Shima Abiati •Kashan University of Medical Sciences

A collaborative research portfolio in cognitive neuroscience and clinical psychology presented at the 1st Kashan Neuroscience Workshop for Students. Investigated the neural substrates of creativity, hypnotic analgesia, and psychological models of learning, earning 2nd Place.

Cognitive Neuroscience Neuroanatomy Divergent Thinking Hypnotic Analgesia Limbic System
Oct 2015 – Feb 2016 Other
Khawrazmi Research Award

Three Centuries in Pursuit of the Theory of Everything

Maryam Rezaee (Independent Researcher) •National Organization for Development of Exceptional Talents

A comprehensive theoretical monograph surveying the historical and mathematical milestones toward a unified field theory in physics, from Newtonian mechanics to contemporary quantum field theories and superstring models. Awarded 2nd Place in the Khawrazmi Youth Award.

Theoretical Physics Unified Field Theory General Relativity Electroweak Unification Khawrazmi Award

Not the End

Forever Chasing Work

Research Interests

Explainable AI Cognitive AI LLM Reasoning Neurosymbolic AI