Agential approach to quantum thermodynamics

Ruo Cheng Huang, Paul M. Riechers, Varun Narasimhachar, Mile Gu

Keywords: Computational Mechanics, quantum information, quantum thermodynamics

TL;DR: In this paper, we discuss the agential framework for quantum thermodynamics, namely how an agent equipped with classical memory can extract useful work from temporal quantum processes.

Abstract: Interacting with dynamic, uncertain environments requires AI agents to perform predictive inference under informational and physical constraints. We present a prototype thermodynamic agent—a pattern engine—that extracts useful work from a non-Markovian quantum process, leveraging principles from computational mechanics. The environment is modeled as a classical hidden Markov model (HMM) with quantum outputs, and the agent maintains an internal belief state that synchronizes to the latent dynamics of this process. Critically, the agent’s performance is governed by the meta-dynamics of its belief updates, which capture the interaction between the environment’s hidden evolution and the agent’s internal representation. We demonstrate that belief-informed policies consistently outperform memoryless and classical strategies, and identify phase transitions in performance linked to bifurcations in the belief dynamics. These findings suggest that alignment failures can emerge not solely from policy design flaws, but from structural limitations in the agent’s ability to accurately track latent information embedded in the data.

Confirmation: I confirm that I and my co-authors have read the policies are releasing our work under a CC-BY 4.0 license.

Submission Number: 4