Research

Understanding how LLMs memorize — whether in weights, context, or external stores — and using that knowledge to build faster inference, principled forgetting, and better training data through synthetic data curation.

Research Directions

Primary Focus

AI Memorization

How LLMs store and recall knowledge — whether in-parameter (weights), in-context (KV cache), or via external retrieval. Research on forgetting, unlearning, KV management, inference optimization, reasoning, and continual learning.

Primary Focus

Synthetic Data

Synthetic data generation, model collapse dynamics, data curation methods, quality filtering, and safety-oriented data pipelines that keep models sharp without the cost of human annotation.

Architecture

Efficient transformer designs, attention optimization, multimodal architectures, and scalable systems. Research on how to maximize model capability per FLOP and unify vision-language processing.

Multi-Agent Systems

How multiple LLM/VLM agents collaborate, debate, verify, and refine outputs. Research on agent orchestration, role specialization, self-play, and multi-agent workflows for complex task decomposition.