My two main interests at the moment are:

Understanding optimization in supervised learning and reinforcement learning from the perspective of optimizing functions (or policies) themselves rather than parameters. Here are papers I co-authored on the topic:
Tighter bounds lead to improved classifiers (ICLR 2017)
An operator view of policy gradient methods (NeurIPS 2020)
A general class of surrogate functions for stable and efficient reinforcement learning (AISTATS 2022 best paper runner-up)
Target-based Surrogates for Stochastic Optimization (ICML 2023)
Decision-Aware Actor-Critic with Function Approximation and Theoretical Guarantees (NeurIPS 2023)

Model adaptation and decentralization of power. How can we create simpler, modular architectures, that can be adapted and reused for other purposes, giving more control to the rightholders (thanks Shakir Mohamed for teaching me that term)? Technical papers I coauthored in that space are:
Multi-Head Adapter Routing for Cross-Task Generalization (in review)
Deep Language Networks: Joint Prompt Training of Stacked LLMs using Variational Inference (in review)