Communications Engineering and Machine Learning Systems

Purbesh Mitra

240-***-**** # ******@***.*** ï linkedin.com/in/purbeshmitra Website Google Scholar Research Interests

6G Communication, Semantic Communication, Information Theory, Machine Learning, Distributed Networks, Wi-fi networks Education

University of Maryland, College Park 2020 – Present Ph.D. in Electrical and Computer Engineering College Park, MD

• Advisor: Prof. Sennur Ulukus

• Focus: Wireless systems, information theory, reinforcement learning, distributed ML, gossip networks Indian Institute of Technology Delhi 2018 – 2020

M.Tech. in Communications Engineering New Delhi, India

• Thesis: Communication in Presence of Non-Gaussian Noise in Wireless Channels Jadavpur University 2014 – 2018

B.E. in Electronics and Telecommunication Engineering Kolkata, India Research Experience

University of Maryland 2021 – Present

Graduate Research Assistant College Park, MD

• Developed Semantic Soft Bootstrapping– a stable self-distillation post-training algorithm from semantic feedback for LLMs without explicit reward function formulation, advancing sample-efficient methods for long context reasoning.

• Developed MOTIF, a novel reinforcement learning framework for multi-iteration reasoning in LLMs that surpasses context window limitations through modular thinking and is trained via verifiable reward mechanism (RLVR), outperforming standard GRPO approach of long context reasoning.

• Designed distributed inference algorithms for edge LLM systems with bounded latency guarantees, enabling collaborative inference that enhances individual model performance in resource-constrained environments.

• Implemented decentralized learning schemes with provable convergence guarantees, achieving O(1) timeliness performance regardless of network size through optimized gossip protocols.

• Developed Gaussian process-based Bayesian optimization for fair timeliness in sparse networks, addressing connectivity heterogeneity in distributed learning systems with gossip communication networks. MediaTek USA May 2024 – August 2024

Research Intern, Software R&D Division San Jose, CA

• Developed deep learning based coordination systems for WiFi-8 multi-station and multi-access point environments.

• Implemented knowledge distillation and memory-efficient neural networks via quantization for edge device deployment. Indian Institute of Technology Delhi 2018 – 2020

Graduate Researcher New Delhi, India

• Developed data-aided channel estimation techniques for non-Gaussian wireless systems with theoretical analysis.

• Analyzed physical layer security metrics using numerical optimization in alpha-stable noise environments. Indian Institute of Science May 2017 – August 2017 Research Intern, Centre for Nano Science and Engineering Bengaluru, India

• Built computer vision pipeline for automated monolayer detection in NEMS device fabrication.

• Designed Arduino-based microscope automation system integrating image processing and mechanical control. Selected Publications (Full list available on Google Scholar) Semantic Soft Bootstrapping: Long Context Reasoning in LLMs without Reinforcement Learning

• P. Mitra, S. Ulukus, submitted to a conference for peer review [Link] MOTIF: Modular Thinking via Reinforcement Fine-tuning in LLMs

• P. Mitra, S. Ulukus, IEEE ICMLA, 2025 [Link]

Distributed Mixture-of-Agents for Edge Inference with Large Language Models

• P. Mitra, P. Kaswan, S. Ulukus, IEEE PIMRC, 2025 [Link] Scale-Robust Timely Asynchronous Decentralized Learning

• P. Mitra, S. Ulukus, IEEE SPAWC, 2024 [Link]

Timely Asynchronous Hierarchical Federated Learning: Age of Convergence

• P. Mitra, S. Ulukus, WiOpt conference: RAWNET workshop, 2023 [Link] A Learning Based Scheme for Fair Timeliness in Sparse Gossip Networks

• P. Mitra, S. Ulukus, IEEE ICMLCN, 2024 [Link]

Technical Skills

Machine Learning: PyTorch, Transformers, Reinforcement Learning, Bayesian Optimization Programming: Python, C, MATLAB

Areas of Expertise: Information theory, distributed networks, machine learning, optimization, reinforcement learning

Contact this candidate