Home 
Research Intern, EPFL NLP Lab
Lausanne, Switzerland
I want to understand how language models reason: the internal mechanisms behind their answers, how those mechanisms emerge through training, and why they sometimes fail.
I am a final-year Computer Engineering student at the University of Tehran and a research intern at EPFL, supervised by Prof. Antoine Bosselut. My current work investigates how distillation configurations affect the reasoning capabilities transferred from teacher to student models.
Previously, at TeIAS, I co-developed PerSo, a benchmark for cognitive and behavioral social intelligence in LLMs. Earlier, I worked on learning-based sound source localization at the University of Tehran’s Human and Robot Interaction Lab.
Our paper PerSo: Investigating Facets of LLMs’ Social Intelligence in Real-World Persuasion Scenarios is now available on arXiv!
I joined the NLP Lab at EPFL as a research intern through Summer@EPFL, supervised by Prof. Antoine Bosselut, to study reasoning and distillation in language models.
I joined the LLM Lab at the Tehran Institute for Advanced Studies to develop a benchmark for social intelligence in language models, supervised by Dr. Mohammad Taher Pilehvar and Dr. Yadollah Yaghoobzadeh.
Our paper, Sound Source Localization in a 3D Multi-Surface Environment Using Learning-Based Methods, was accepted to ICRoM 2024.
I joined the Human and Robot Interaction Lab at the University of Tehran, supervised by Dr. Mehdi Tale Masouleh, to work on learning-based sound source localization.
Evaluates whether LLMs can infer beliefs, emotions, and intentions and use that understanding to act effectively in real-world persuasion scenarios, connecting cognitive and behavioral social intelligence.
12th RSI International Conference on Robotics and Mechatronics, pp. 717–724.
Learning-based region classification reduced localization error from 28.4 to 23.4 cm and runtime from 29.8 to 0.71 seconds.
Aug 2026–present
Studying how distillation configurations shape reasoning and student behavior in language models. My work compares rollout horizons, prefix sources, and KL directions through checkpoint evaluations, and examines answer-format alignment, reasoning loops, and recovery during training.
Jun 2025–Jul 2026
Co-developed PerSo, a benchmark evaluating cognitive and behavioral social intelligence in LLMs through real-world persuasion scenarios. Designed belief, emotion, intent, and persuasion-response tasks, built the dataset with LLM-assisted generation and human validation, and evaluated 15 model configurations across five families.
2024
Developed learning-based 3D sound source localization using audio timing and spectral features. Collected audio at 315 locations and evaluated multi-branch neural networks and region classifiers. Published and presented the work at ICRoM 2024.