Marilyn Rego

Marilyn Rego

she/her

Ph.D. student in Computer Science and Engineering at the University of Michigan.

Research

Diagram from the mrLLM poster: an LLM split across Region A and Region B, with large intermediate activations crossing a bandwidth-limited WAN link.
Aug 2025 – Present · University of Michigan · SRC Research Scholars Program

Communication-Efficient Cross-Region LLM Inference

Building communication-efficient systems for LLM inference across geographically distributed data centers, through hardware–software co-design.

  • NSDI ’26 poster mrLLM: Fast Multi-Region LLM Inference Using Learned Adapters Poster
  • NAIC @ SIGCOMM ’26 Towards Network-Efficient Cross-Regional Inference via Learned Activation Compression Paper
  • Designing learned activation-compression techniques and adapter-based methods to reduce cross-region data movement while preserving model quality.
  • Evaluating partitioning strategies for prefill, decode, and agentic inference workloads across latency, throughput, data movement, and model quality under heterogeneous compute and network constraints.
Listing 1 from the FORGE 2025 paper: a C increment function given to GPT-4o as input, and the output with VeriFast requires/ensures and open/close annotations.
May 2024 – Aug 2024 · Purdue · Summer Undergraduate Research Fellowship

LLMs for Verifiable Specification Generation

Can LLMs write formal specifications that a static verifier (VeriFast) actually accepts?

  • FORGE ’25 Evaluating the Ability of GPT-4o to Generate Verifiable Specifications in VeriFast Paper
  • ISSRE ’26 An Empirical Study of LLM-Generated Specifications for VeriFast Details
  • Evaluated LLMs on generating separation logic specifications for VeriFast, from baseline prompting to chain-of-thought, prompt chaining, and few-shot prompting.
  • Categorized the VeriFast benchmarks and built tailored input datasets for systematic evaluation.
  • Designed a qualitative analysis framework for output correctness, syntactic validity, and model failure modes.
Diagram from the SpliDT poster: a decision tree partitioned into subtrees p1–p4, each using its own window of packets w1–w4.
SRC TECHCON ’25 Jan 2024 – Aug 2025 · Purdue · NextGArch Lab

SpliDT: Partitioned Decision Trees for Line-Rate ML Inference

Detecting malicious network traffic with ML models that run at line rate on programmable switches.

Poster

  • Implemented and benchmarked one-class SVMs, random forests, and isolation forests for payload-level anomaly detection, establishing baselines for accuracy, cost, and scalability.
  • Analyzed how packet-, flow-, and connection-level payload features trade off prediction accuracy against computational overhead.
  • Explored transformer-based models to automate payload feature selection, comparing them against traditional methods integrated with SNORT.
May 2023 – Aug 2023 · Purdue · Department of Chemistry

Deep Learning for Drug Discovery

A deep-learning pipeline that predicts 3D molecular structures and properties to study crystal growth.

  • Implemented a Python deep-learning pipeline that predicts 3D molecular structures and properties from user-input parameters.
  • Applied ML methods to model crystal growth mechanisms and improve prediction accuracy.