Selected Highlights
A Concept-Level Energy-Based Framework for Interpreting Black-Box LLM Responses
Tackles the critical interpretability challenge of closed-API LLMs. We propose a model-agnostic framework that uses an energy model to score prompt-response consistency. This energy trains a lightweight interpreter into an efficient, standalone tool that explains LLM outputs by quantifying prompt influence—all without further API calls.
View Proof-of-ConceptConcept-Based Interpretability for RAG Systems
Investigating the use of Concept Bottleneck Models (CBMs) to explain internal generation processes as a Research Mentor in the Trustworthy and Generative ML Lab. This work aims to enhance the transparency and traceability of RAG outputs by extracting and tracing concepts using Concept Activation Vectors (CAV) and ACE.
See Details (Coming Soon)A Neuro-Symbolic Architecture for Translating Literary Formulas into Coherent Narrative Generation
Proposes a “Director-Actor” architecture to solve the “wandering plot” problem in LLMs. We operationalize literary theory into computational algorithms—modeling dramatic arcs via Signal Processing, Gated FSMs, and A* Search—framing narrative as an Active Inference optimization task.
View PaperNeurosymbolic VQA Program Generator
An exploration of neurosymbolic VQA (Johnson et al., 2017) on the CLEVR dataset. This project implements and compares three distinct strategies for translating natural language questions into executable symbolic programs: Supervised Learning (LSTM/Transformer), Reinforcement Learning (REINFORCE), and In-Context Learning (LLM).
View on GitHubCore Technologies
Current Focus
M.Sc. in Computer Science
Sharif University of Technology
Master’s Thesis
Currently working on my Master’s Thesis on “Interpretability in Generative Models: Investigating the Mechanisms Behind Output Generation in Large Language Models.”
Research Exploration
Conducting independent research into Neurosymbolic Architectures and LLM Reasoning with a focus on integrating principles from Cognitive Science to build human-like systems.
System Status: Expanding Archives
Detailed pages for About, Research, Experience, Projects, and Talks are currently under active development. For now, please refer to the Curriculum Vitae as an overview.