About Me

I am Akshat Valse, an undergraduate in the Honors Program at Iowa State University, where I am working toward four degrees: B.S. Statistics, B.S. Mathematics, B.A. Computer Science and B.S. Data Science. I started in August 2024 and expect to graduate in May 2028. I am applying to PhD programs in machine learning and applied probability for Fall 2028 admission.

In Fall 2026 I am taking four PhD-level courses: measure-theoretic probability, and advanced topics in machine learning, artificial intelligence and high-performance computing. I run jobs with SLURM on Iowa State's Nova cluster, and serve models with SGLang and vLLM on NVIDIA A100 and H200 GPUs. The details are in my CV, also available as a PDF.

My research uses probability and numerical analysis to study how machine learning systems behave. I work on six projects with six advisors in Computer Science and Mathematics at Iowa State, and they fall into three groups.

• LLM serving and agents: Inside the Trap, my preprint with Ali Jannesari of the SwAPP Lab (arXiv:2611.08902), explains why a burst of agent traffic can leave an LLM serving engine stuck at a load it handled before, and how a session cap gets it out. Agent Reliability, advised by Yang Li, asks when restarting a stalled agent is worth more than waiting, and how to estimate the probability of a failure too rare for direct simulation to reach. It is in experimentation and development.
• Sampling and numerical analysis: Langevin Sampling, with Dr. David Herzog, checks Langevin samplers against answers known in closed form; that part of the project and its code are public. A manuscript with him on transport-coefficient estimation for molecular systems is drafted and in his review. Learned PDE Solvers, with Jue Yan and Mengyuan Yang, asks whether you can tell from a trained cell-average neural network solver alone, before running a long simulation, whether its rollout will stay stable. We are adapting von Neumann stability analysis to these solvers, and the manuscript is in preparation.
• Rare events and uncertainty: Rare Events, with Ruoyu Wu, uses large-deviation theory to guide importance sampling for rare events in randomized load balancing (the mean-field join-the-shortest-queue model), such as a queue climbing to a high level within a time window. It is paused during my advisor’s sabbatical. Early-Feedback Calibration, advised by Hongyang Gao, is about volatility forecasts. A forecast of a cumulative quantity is scored when its horizon ends, yet a miss can become certain before then; I am developing a calibration method that uses this early, one-sided feedback. The project is in its experimentation and development phase.

When I am not running experiments, I fly. I hold an FAA private pilot certificate with a high-performance endorsement, and I volunteer with Pilots N Paws on its animal-rescue flights, as a pilot and as a recruitment advisor.

• Nov. 2026. Preprint with Ali Jannesari (SwAPP Lab) posted: Inside the Trap: A Renewal Law for Prefix-Cache Collapse in LLM Agent Serving, arXiv:2611.08902. The code, all real-engine data and every registration are public.
• Aug. 2026. Started three projects: metastable failures in prefix-cached LLM agent serving with Ali Jannesari, early-feedback calibration of volatility forecasts with Hongyang Gao, and reliability of tool-using LLM agents with Yang Li.
• May 2026. Joined David Herzog on Langevin sampling and diffusion estimation (funded by NSF award DMS-2246491) and Jue Yan on the stability of cell-average neural networks for PDEs (funded by the NSF D4 NRT; joint with Mengyuan Yang).
• Nov. 2024. Undergraduate Research Fellow with Ruoyu Wu on large deviations and rare events in queueing systems.
• Aug. 2024. Entered Iowa State University's Honors Program, working toward four degrees.

• Dean's High Impact Undergraduate Research Award, 2026
• Schillmoeller Family Scholarship in Statistics
• Mead Fund Award in Statistics
• James L. Cornette Mathematics and Interdisciplinary Studies Award
• Dean's List, every semester