Open and completed research requests

We propose a reinforcement learning methodology that combines intrinsic novelty-driven exploration with explicit goal conditioning and extrinsic reward optimization. The core motivation is to avoid myopic exploitation while preserving path optimality for objective goals. The approach uses a dual-value design: one stream for task-optimal return and one for curiosity, with adaptive weighting that decays as goal confidence improves. We expect this to improve sample efficiency, robustness, and discovery of better trajectories in sparse- and deceptive-reward settings. The impact is a practical framework for agents that remain purpose-driven while retaining productive curiosity.

Continual learning systems must remain plastic enough to learn new tasks while stable enough to avoid catastrophic forgetting, but this tradeoff is increasingly constrained by memory behavior rather than compute alone. This project adopts a unified data-access perspective in which deterministic retrieval and stochastic sampling are treated within one framework, so memory support for both adaptation and retention can be analyzed jointly. We will design and evaluate memory mechanisms that explicitly manage stochastic demand, entropy pressure, and replay/update efficiency across task sequences. Expected impact is a clearer systems-level blueprint for continual learning stacks that improve retention and adaptation while remaining robust and efficient on realistic hardware.

This project targets a key safeguards challenge: identifying covert fissile production pathways in nuclear fusion reactor settings using antineutrino observations. We will study whether novel material configurations (blankets, moderators, shielding, and structural composites) create measurable antineutrino-spectrum or time-profile signatures that improve detectability. The approach combines physics-informed simulation of reactor scenarios with statistical detection models and material sensitivity analysis. We expect to produce a ranked set of material candidates and monitoring strategies that improve early-warning capability while quantifying uncertainty and false-alarm risk.

We introduce SymbolicAI, a versatile and modular framework employing a logic-based approach to concept learning and flow management in generative processes. SymbolicAI enables the seamless integration of generative models with a diverse range of solvers by treating large language models (LLMs) as semantic parsers that execute tasks based on both natural and formal language instructions, thus bridging the gap between symbolic reasoning and generative AI. We leverage probabilistic programming principles to tackle complex tasks, and utilize differentiable and classical programming paradigms with their respective strengths. The framework introduces a set of polymorphic, compositional, and self-referential operations for multi-modal data that connects multi-step generative processes and aligns their outputs with user objectives in complex workflows. As a result, we can transition between the capabilities of various foundation models with in-context learning capabilities and specialized, fine-tuned models or solvers proficient in addressing specific problems. Through these operations based on in-context learning our framework enables the creation and evaluation of explainable computational graphs. Finally, we introduce a quality measure and its empirical score for evaluating these computational graphs, and propose a benchmark that compares various state-of-the-art LLMs across a set of complex workflows. We refer to the empirical score as the "Vector Embedding for Relational Trajectory Evaluation through Cross-similarity", or VERTEX score for short. The framework codebase and benchmark are linked below.

Estimate achievable X-gate speeds and fidelities for a Yb-171 nuclear qubit using Raman transfer via a long-lived inner-shell J=2 intermediate (with 1 W/cm^2 available), limited by hyperfine splitting versus intermediate lifetime, and report gate-time vs target-infidelity tradeoffs (10^-2-10^-5).

We propose the problem of designing a dynamic, continuously-adaptive activation function that embeds continual-learning inductive biases directly into neuron nonlinearity. The target activation should trade off plasticity versus stability so networks can incrementally integrate concepts from multiple domains without catastrophic forgetting, while remaining scalable (avoiding saturation/dead neurons) and self-stabilizing (preserving bounded mean/variance across layers). Crucially, this mechanism must not rely on predefined task counts, explicit task identities, or brittle masking hacks; instead it should fold related concepts into shared representations and enable layer-wise progressive knowledge integration and reuse so that higher-level abstractions and deductive reasoning can emerge.