Group
Arkansas Machine Learning Group
Learning to decide in complex worlds.
The Arkansas Machine Learning Group develops machine learning methods for adaptive decision-making in complex systems. Our research focuses on sequential decision-making processes, search, exploration, and planning for language models, and artificial intelligence for scientific and operational decision-making.
Research agenda
Across these areas, we ask a common question: how can a learning system use scarce feedback to choose what to try, where to search, and how to act next?
Sequential Decision-Making Processes
We develop sample-efficient learning and optimization methods for complex decision-making problems, especially those involving structured inputs, outputs, actions, and constraints. Our interests include Bayesian optimization, active learning, and uncertainty-aware sequential decision-making over graphs, sets, sequences, combinatorial spaces, and other structured domains.
Recent examples
- Density-ratio Bayesian optimization with semi-supervised learning ICML 2025
- Noise-adaptive confidence sets for linear bandits and Bayesian optimization ICML 2024
- Combinatorial Bayesian optimization through continuous-space mappings UAI 2022
- Budget-aware sequential construction under combinatorial constraints TMLR 2024
Search, Exploration, and Planning for Language Models
We study how language models can search, explore, and plan more effectively through feedback and structured decision processes. Our work aims to improve the reliability and efficiency of language-model systems by developing methods for test-time search, tool use, agentic planning, and reward modeling.
Recent examples
- Transformers navigating unknown search spaces from bandit feedback TMLR 2026 with J2C Certification
- ReJump: analyzing exploration, backtracking, and verification in reasoning ICML 2026
- TAPE: tool-guided adaptive planning with constrained execution ICLR AAIW 2026
- VersaPRM: process reward modeling beyond mathematical reasoning ICML 2025 Oral
Artificial Intelligence for Scientific and Operational Decision-Making
We build artificial intelligence systems that support reliable decisions in scientific, engineering, and operational domains. We are especially interested in problems where learning systems must reason over structured data, uncertain objectives, limited feedback, and real-world constraints, with applications in scientific discovery, engineering design, logistics, and supply chains.
Recent examples
- CHROMA: natural-language inverse design of structural coloration NeurIPS AI4Mat 2025
- Datasets and benchmarks for nanophotonic structure and parametric design simulations NeurIPS 2023
- Multi-fidelity, multi-objective Bayesian optimization for nanophotonic design Digital Discovery 2024
- JaxLayerLumos: differentiable simulation for multilayer structures JOSS 2025
Members
To be announced.
To be announced.
Join the group
Prospective students should send a CV and a short research statement to arkansas.mlg+application@gmail.com. We also welcome conversations with research collaborators working across machine learning, science, engineering, and operations.