Evan Chen (Po-Yu Chen)

PhD Candidate, Elmore Family School of Electrical and Computer Engineering, Purdue University, West Lafayette, IN · chen4388@purdue.edu · CV

I am an ECE PhD candidate at Purdue University, advised by Professor Christopher G. Brinton, with an expected graduation in December 2026. My research focuses on scalable and reliable distributed AI systems, particularly LLM-based multi-agent systems and collaboration between on-device and cloud language models.

I develop communication-aware coordination and consistency mechanisms for heterogeneous agents, along with reinforcement-learning and training-free methods that let local language models decide when cloud assistance is worthwhile under resource budgets.

My broader work spans federated and fog learning, distributed optimization, and differential privacy. I am interested in turning principled learning and networking methods into dependable AI systems that operate under real-world communication, computation, privacy, and reliability constraints.

Multi-Agent Systems

Scalable coordination and consistency for heterogeneous LLM agents.

On-Device + Cloud LLMs

Resource-aware routing and RL post-training for collaborative reasoning.

Distributed + Private Learning

Federated optimization across communication, trust, and privacy constraints.

News

  1. I will join Nokia in Sunnyvale as a Research Intern in Fall 2026.
  2. I joined Microsoft Research, Redmond as a Research Intern working on security ML systems.
  3. Our paper Bridging On-Device and Cloud LLMs for Collaborative Reasoning was accepted to ICML 2026.
  4. Our hierarchical gradient-tracking work was published in IEEE/ACM Transactions on Networking.
  5. I advanced to PhD candidacy at Purdue University.

Publications

Preprint · 2026

A Survey on Jev: Prior Research, Adaptation Costs, and Evidence of System One Models

A survey of typed decision models such as Jev, which return probabilities over a caller-defined answer set instead of generated text: their roots in LLM routing and zero-shot classification, the cost of adapting them to new models, decisions, and constraints, and what current evidence supports.

Paper
  • Large Language Models
  • Typed Decision Models
  • LLM Routing
  • Survey

Preprint · 2026

Routing Without Training: Controllable-Ratio LLM Offloading via Reliability Gating

CARGO uses inference-time response agreement, Bayesian early stopping, and lightweight calibration to route requests between local and cloud LLMs at a controllable collaboration ratio—without training an additional router.

Paper
  • Large Language Models
  • On-device AI
  • Cloud Offloading
  • Training-free Routing

NeurIPS 2026 · Under Review

Building Scalable Multi-Agent LLM Systems under Heterogeneous Communication, Capability, and Tasks

A communication-aware framework for coordinating heterogeneous LLM agents across varying capabilities, tasks, and network constraints.

  • Multi-Agent Systems
  • Large Language Models
  • Communication

NeurIPS 2026 · Under Review

Joint Continual Learning of Local Language Models and Cloud Offloading Decisions with Budget Constraints

DA-GRPO jointly improves a local language model and its cloud-collaboration behavior while controlling assistance budgets and reducing forgetting across changing tasks.

Paper
  • Large Language Models
  • RL Post-Training
  • Continual Learning
  • Cloud Offloading

ICML 2026

Bridging On-Device and Cloud LLMs for Collaborative Reasoning: A Unified Methodology for Local Routing and Post-Training

A unified approach that teaches on-device LLMs to reason locally and invoke cloud assistance selectively through reinforcement-learning-based post-training.

Paper
  • Large Language Models
  • On-device AI
  • Reinforcement Learning
  • Collaborative Reasoning

Skills

Languages, operating systems & tools

  • Python
  • PyTorch
  • Linux
  • LaTeX
  • MATLAB
  • C/C++
  • Swift
  • Java
  • NVIDIA FLARE
  • ExecuTorch

Research interests

  • Multi-Agent Systems
  • Large Language Models
  • RL-based Post-Training
  • On-device Model Inference
  • Reinforcement Learning
  • Optimization
  • Communications and Networking
  • Differential Privacy
  • Federated Learning

Experience

Research Intern

Nokia

Research internship in Sunnyvale, California.

Research Intern

Microsoft Research

Improved the stability and performance of a production-scale security ML system and developed configurable agent-based simulations for realistic enterprise interaction traces.

Academic Services

Reviewer

  • Conference on Neural Information Processing Systems (NeurIPS)
  • International Conference on Machine Learning (ICML)
  • International Conference on Learning Representations (ICLR)
  • IEEE International Conference on Computer Communications (INFOCOM)
  • IEEE Transactions on Mobile Computing (TMC)
  • IEEE/ACM Transactions on Networking (ToN)
  • IEEE/ACM Transactions on Parallel and Distributed Systems (TPDS)

Presenter

  • “Taming Subnet-Drift in D2D-Enabled Fog Learning: A Hierarchical Gradient Tracking Approach,” INFOCOM, Vancouver, Canada, May 23, 2024.

Education

Purdue University West Lafayette, IN, USA

Doctor of Philosophy

Elmore Family School of Electrical and Computer Engineering
PhD candidate advised by Professor Christopher G. Brinton · Expected December 2026