LLM + RL for agents
Training and evaluating agents that must act, not just answer, by bridging language understanding with sequential decision-making.
policy optimization for language agents • reward design • long-horizon tasks • evaluation
The Ohio State University · Computer Science
Oliver Proudf oo t
PhD student in AI
Research at the boundary of large language models (LLMs) and reinforcement learning (RL), with an emphasis on building systems that reliably reason, decide, and improve with experience.
I work on methods that connect language-based reasoning with decision-making and learning from interaction, aiming to make LLM-driven agents more robust, controllable, and effective on real tasks.
Advisor: Andrew Perrault
LLMs · RL · agents · alignment/safety · optimization
About
I am a PhD student focused on AI, working at the interface of LLMs and RL. My research explores two directions: applying RL-style feedback and objectives to shape LLM behavior, and using LLMs as components inside RL agents for reasoning, planning, tool use, and action selection.
The central theme is turning language models into reliable decision-makers, not just text generators, through clear objectives, strong empirical evaluation, and reproducible system design.
Research
Reliable reasoning and learning signals
Methods for making improvements measurable and verifiable with explicit objectives, robust benchmarks, and careful feedback design.
process supervision • calibration • objective alignment • robust scoring
Systems + experiments
Building reproducible training and evaluation pipelines for large-scale experimentation with practical rigor.
distributed runs • reproducibility • toolchain reliability • experiment ops
Selected Projects
See more
ND Optimizer Arena
Interactive optimization arena and benchmark environment for studying optimizer behavior.
Decomposed Reasoning
Framework for structuring language-agent reasoning and evaluating reliability across staged decomposition pipelines.
Agent Evaluation Suite
Standalone benchmarking and average@k style analyses for comparing model reasoning behavior under controlled settings.