Research

My research focuses on building AI systems that can adapt, remember, evolve, and transfer across changing tasks, domains, modalities, and scientific problems. Some of these directions are reflected in my publications; others are active interests that I am currently exploring.

Current Interests

Agentic Memory and Evolving Agent Systems

I am interested in how agents can maintain useful long-term memory, retrieve relevant experiences, and update their behavior through interaction. This includes memory mechanisms for agentic workflows, multi-agent collaboration, self-evolving systems, and methods that allow agents to improve their strategies without relying only on static training data.

Transfer, Meta, and Continual Learning

I study how models reuse knowledge when the environment changes. This direction includes transfer learning, meta-learning, domain adaptation, gradual distribution shifts, continual learning, and lifelong learning. I care about both the algorithmic side, such as optimization and transport views, and the practical side, such as preventing forgetting while adapting to new tasks.

Efficient Adaptation of Large Models

I work on parameter-efficient and resource-aware adaptation for large models, including LLMs, diffusion models, and multimodal foundation models. I am especially interested in low-rank adaptation, shared or modular adaptation, adaptive capacity allocation, distillation, and training/inference-efficient methods that make large models easier to specialize.

New Architectures for Generative Models

I am broadly interested in new generative model architectures and learning paradigms beyond standard pipelines. This includes diffusion language models, concept-level models, recurrent generative models, spiking neural networks, and other architectures that may improve generation, reasoning, memory, controllability, or efficiency.

AI for Science

I also care about applying AI to scientific discovery and structured reasoning. Current interests include mathematics, physics, cosmology, neuroscience, and other domains where learning systems can help model complex phenomena, discover reusable structure, or support reasoning over scientific knowledge.

For a detailed publication list, see Publications.