The Department of Systems and Information Engineering at the University of Virginia seeks postdoctoral research associates to develop statistical and algorithmic foundations for reinforcement learning from human feedback (RLHF) and epistemic control of large language models (LLMs). The project will study how human preferences and other forms of feedback can be used to train LLMs that reason reliably, recognize uncertainty, and adapt their behavior to the task and the user.

The position is expected to begin on January 1, 2027. The initial appointment will be for one year, with the possibility of renewal for an additional year, contingent on satisfactory performance and the availability of funding.

Research Program

Although RLHF and related methods, including direct preference optimization, are now widely used, their statistical properties remain only partially understood. Human feedback is noisy, heterogeneous, context dependent, and shaped by the process through which data are collected. The postdoctoral researcher will investigate questions involving identifiability, sample complexity, generalization, uncertainty quantification, reward misspecification, and the propagation of estimation error from preference models to learned policies.

The work may also develop adaptive methods for collecting human feedback more efficiently.

A complementary research direction concerns epistemic control: treating an LLM as a controlled reasoning system rather than as a one-shot response generator. A high-level controller may direct the model to decompose a problem, generate alternative hypotheses, retrieve information, verify evidence, check consistency, request clarification, calibrate confidence, or defer judgment. The project will formulate these choices as a hierarchical decision problem in which an epistemic controller selects reasoning actions that the LLM executes through language generation, structured reasoning, or tool use.

The postdoctoral researcher will contribute to theory, algorithms, and empirical evaluation. Possible research outcomes include finite-sample guarantees for preference-based estimators, uncertainty-aware reward modeling, off-policy evaluation methods, adaptive experimental designs, and algorithms for epistemic control under partial observability. Empirical studies may use open-source LLMs and benchmark tasks in reasoning, scientific question answering, code review, tutoring, or decision support.

Responsibilities

The successful candidate will be expected to:

• Develop theoretical and computational methods for RLHF, preference learning, and epistemic control;

• Establish statistical or algorithmic guarantees where appropriate;

• Design and implement empirical evaluations using modern machine-learning frameworks and open- source LLMs;

• Prepare research papers for publication in leading machine-learning, artificial-intelligence, statistics, operations research, or systems venues;

• Collaborate with faculty, graduate students, and other project researchers

• Contribute to the intellectual development of a broader research program on reliable and human-centered AI.

Qualifications

A Ph.D. in electrical engineering, computer science, statistics, applied mathematics, operations research, systems engineering, or a closely related field by the appointment start date.

Preferred Qualifications:

• Strong background in theoretical machine learning, statistical learning theory, optimization, or reinforcement learning;

• Experience establishing theoretical guarantees for deep-learning or other high-dimensional statistical models;

• Research on efficient training or inference for large neural networks, including mixture-of-experts models, pruning, quantization, or related methods;

• Experience implementing and evaluating large models using frameworks such as PyTorch, JAX, or TensorFlow;

• Interest in extending theoretical and computational expertise toward RLHF, human-centered AI, LLM alignment, and reliable reasoning.

•Experience with large language models, mechanistic interpretability, preference learning, inverse reinforcement learning, human-feedback data, or human-subject experimentation is desirable but not required.

•The position is especially well suited for a researcher interested in connecting rigorous machine-learning theory with the development of efficient, transparent, and reliable LLM systems.

TO APPLY

Apply at https://uva.wd1.myworkdayjobs.com/UVAJobs and search for R0085977. Attach a cover letter, a detailed curriculum vitae, and contact information for three references. Please note that multiple documents can be uploaded in the CV box.

APPLICATION DEADLINE: Review of applications will begin on August 16, 2026, and the posting will remain open until filled. The University will perform background checks on all new hires prior to employment.

Estimated Salary range is $62,000 - $67,000, commensurate with experience.

For questions regarding this position, please contact Matt Sinclair at bkb9bf@virginia.edu.

For questions regarding the application process, contact Rich Haverstrom, Academic Recruiter, at rkh6j@virginia.edu.

For more information on the benefits available to postdoctoral associates at UVA, visit postdoc.virginia.edu and hr.virginia.edu/benefits.

The University of Virginia is an equal opportunity employer. All interested persons are encouraged to apply, including veterans and individuals with disabilities. Learn more about UVA’s commitment to non-discrimination and equal opportunity employment.