Research portfolio

Human feedback, intelligent agents, and the behavior of learning systems.

I study how models learn from structured or imperfect feedback, how architecture and capacity shape generalization, and how to evaluate whether learned behavior is useful and reliable.

Core question

How can machines learn useful, generalizable behavior from human judgments and constrained data?

Research process

Problem formulation, algorithm design, implementation, controlled experimentation, analysis, and publication.

Technical scope

RLHF, reward modeling, reinforcement learning, compact language models, agents, evaluation, and autonomous control.

2025 ICML 2025 · TTODLer-FM Workshop · Oral

Too Big to Think: Capacity, Memorization, and Generalization in Pre-Trained Transformers

Joshua Barron, Devin White

Controlled character-level pretraining experiments examine how model capacity changes the balance between arithmetic extrapolation and factual memorization. The results identify a narrow generalization window that collapses when factual knowledge is introduced.

My contribution: Advised on research framing, interpretation of experimental results, and communication of the capacity–generalization findings.

LLMs Small models Generalization
2025 ICML 2025 · Models of Human Feedback Workshop

Multi-Task Reward Learning from Human Ratings

Mingkang Wu, Devin White, Evelyn Rose, Vernon Lawhern, Nicholas Waytowich, Yongcan Cao

Combines rating classification and reward regression with learnable uncertainty weights, providing a more label-efficient approach to reward modeling from human ratings.

My contribution: Contributed to problem formulation, investigation and interpretation of results, research direction, and manuscript development.

RLHF Ratings Reward modeling
2025 NeurIPS 2025 · ARLET Workshop

Human-Inspired Multi-Level Reinforcement Learning

Mingkang Wu, Devin White, Vernon Lawhern, Nicholas Waytowich, Yongcan Cao

Uses rated performance levels to guide both reward learning and policy improvement, including information from failed and lower-performing experiences rather than learning only from successful behavior.

My contribution: Contributed to problem definition, investigation of results, research direction, next-step decisions, and manuscript development for the expanded multi-level framework.

RLHF Ratings Policy learning
2025 AAAI 2025 · CAIHU Bridge

Performance Optimization of Ratings-Based Reinforcement Learning

Evelyn Rose, Devin White, Mingkang Wu, Vernon Lawhern, Nicholas Waytowich, Yongcan Cao

Systematically investigates design choices that affect the consistency and performance of Rating-Based Reinforcement Learning, including rating boundaries, smoothing, and the number of rating classes.

My contribution: Helped define the research questions, investigate consistency and performance tradeoffs, determine next experiments, and develop the manuscript.

RLHF Ratings Evaluation
2025 AAAI 2025 · KnowFM Workshop

Atari-GPT: Benchmarking Multimodal Large Language Models as Low-Level Policies in Atari Games

Nicholas R. Waytowich, Devin White, MD Sunbeam, Vinicius G. Goecks

Introduces a benchmark for evaluating multimodal language models as low-level controllers in Atari-style environments, exposing limitations in visual grounding, action consistency, and real-time interaction.

My contribution: Helped define the research direction; developed code and environments; ran experiments; and contributed to analysis and writing.

LLMs Agents Benchmarking
2024 AAAI 2024

Rating-Based Reinforcement Learning

Devin White, Mingkang Wu, Ellen Novoseller, Vernon J. Lawhern, Nicholas Waytowich, Yongcan Cao

Proposes learning reward functions from direct ratings of individual trajectories rather than pairwise preferences, supported by synthetic-feedback experiments and an IRB-approved human-subject study.

My contribution: Helped originate the research direction and led algorithm design, implementation, experimental planning, analysis, and publication.

RLHF Ratings Human study
2024 AIAA SciTech 2024 · GNC

Deep Reinforcement Learning-based Optimal Time-constrained Intercept Guidance

Abhinav Sinha, Devin White, Yongcan Cao

Uses proximal policy optimization to select guidance-law design parameters for energy-efficient rendezvous under a specified arrival-time constraint.

My contribution: Contributed to reinforcement-learning system development and experimental analysis for the autonomous guidance problem.

Reinforcement learning Control Autonomy
2023 ICML 2023 · Workshop

Rating-based Reinforcement Learning

Devin White, Mingkang Wu, Ellen Novoseller, Vernon Lawhern, Nicholas Waytowich, Yongcan Cao

Early workshop presentation of the direct-rating framework that later developed into the full AAAI 2024 Rating-Based Reinforcement Learning paper.

My contribution: Led early method development and evaluation that established the foundation for the later full-conference work.

RLHF Ratings Reinforcement learning

How I work

Research ownership across the full lifecycle.

01

Frame the problem

Identify the decision, assumption, or limitation worth testing and define what evidence would change the conclusion.

02

Build the system

Translate the idea into algorithms, training pipelines, environments, data processing, and reproducible experiments.

03

Evaluate deeply

Compare baselines, inspect failure modes, test robustness, and use metrics that reflect the actual research question.

04

Communicate clearly

Turn results into papers, presentations, open-source implementations, and actionable next steps for collaborators.

Master's thesis · UTSA · 2023

Reinforcement Learning From Human Ratings

A unified account of the development and evaluation of Rating-Based Reinforcement Learning, including algorithm design, synthetic experiments, and human feedback.

View thesis ↗