Quantum Transformers represent a new research direction at the convergence of quantum computing and classical Transformer architectures, aiming to overcome fundamental limitations of purely classical learning systems in real-world environments. Classical Transformers excel under controlled conditions, yet their performance and trustworthiness often degrade in the presence of noise, drift, missing data, domain shifts, and strict resource constraints. At the same time, quantum computing introduces powerful primitives, such as quantum feature spaces, quantum optimization, and simulation-inspired dynamics, while also facing practical challenges, including noisy hardware, limited qubit availability, and sampling overhead. This research explores quantum-aware Transformer methodologies and deployment-first engineering practices that strategically combine quantum and classical capabilities to enable robust, scalable, and reproducible learning systems. It advances three tightly integrated thrusts: scalable representation learning and global mixing, developing efficient long-context models for multivariate time series and multimodal sensing through compute-aware alternatives to standard attention, context-aware fusion of heterogeneous sensors, and physics or measurement informed embeddings that enhance stability and interpretability; quantum-enhanced learning, reasoning, and optimization, investigating hybrid quantum, classical pipelines that leverage quantum feature mapping, quantum optimization, and simulation-inspired modules to accelerate difficult inference and search tasks, improve generalization in high-dimensional regimes, and reduce labeling requirements where possible; and trust, verification, and deployment-first evaluation, establishing rigorous protocols that include leakage-resistant data splits, stress testing under noise, missingness, and adversarial corruption, calibrated uncertainty estimation, and operational profiling of latency, memory, and energy proxies alongside parameter-efficient adaptation. Collectively, this work aims to move Quantum Transformers from conceptual promise toward validated, trustworthy systems capable of delivering measurable impact in real-world scientific and engineering applications.

Reinforcement learning (RL) underpins many modern decision-making systems, from robotics and resource scheduling to adaptive control and autonomous agents. However, as problem complexity grows, classical RL can become bottlenecked by high-dimensional state spaces, expensive policy evaluation, and slow exploration in large action spaces. This project investigates Quantum Reinforcement Learning (QRL) as a principled route to accelerate or enhance RL pipelines by leveraging quantum representations and quantum-assisted optimization, while remaining compatible with near-term (NISQ) hardware through hybrid quantum–classical training loops.

We will develop and benchmark quantum-enhanced RL agents that integrate parameterized quantum circuits (PQCs) as policy and/or value-function approximators, and explore quantum subroutines for components such as feature embedding, policy search, and exploration. The work will emphasize theoretical soundness, clearly specifying what is quantum, what remains classical, and how learning signals (returns, advantages, gradients) are estimated under realistic constraints such as finite sampling (“shots”) and device noise. We will evaluate performance using reproducible benchmarks and establish practical guidelines for when quantum components can provide measurable benefits (e.g., sample efficiency, expressivity under parameter constraints, robustness, or computational advantages in specific regimes).

The project aims to deliver (i) a coherent QRL framework suitable for real devices, (ii) open, reusable benchmarking code and protocols, and (iii) a set of well-characterized use cases, such as control, scheduling, and resource allocation, where QRL is most likely to offer value for next-generation intelligent systems.