Ryan Lagasse

Image from another paper, used as a stand-in.

Targeted LLM Steering: Mitigating Side Effects with Selective Feature Control

Improving activation steering vectors by formalizing constructive and destructive features in steering vectors to revert entangled features with negative side effects to their unsteered states. Accepted to RAISE: Winter Expo

Extending Finetuning Scaling Laws for LLMs

Exploring extending scaling laws to comprehensively select LLM sizes for small datasets based on token length and number of examples improving on existing work. Submitted to ICLR 2025 Workshop

Hybrid Quantum Algorithms for N-Body Simulations

Exploring hybrid quantum algorithms to accelerate n-body simulations. Accepted to QCNC 2025!

LLM Alignment Research

Proposing RLOF, a scalable fine-tuning method for training LLMs with small datasets. Targeted for NEURIPS 2025

Persistent Refinement in Transformers

Improving chat bias benchmarks with innovative transformer steering methods. Targeted for NEURIPS 2025

UAV-UGV Teaming

Led research for a very successful first of its kind autonomous robotics teaming demonstration at EDGE24 Conference

Electric Grid Software Prediction Modeling

Led development on software failures prediction model for over 7 million Aclara electric meters across the United States with 99% accuracy (AUC)

Detection of Ventricular Arrhythmias

Optimized CNN-based classifier, winning top accuracy at TinyML track at ICAD.