Zizhao Hu
AI Researcher · PhD Student at USC · GLAMOUR Lab · MINDS · MOVE Fellow
“Understanding how LLMs remember and forget—then using that knowledge to build faster, leaner inference and better training data.”
— Zizhao Hu, PhD Student at USC · GLAMOUR Lab & MINDS Group
LLM Memorization
How LLMs remember & forget — unlearning, KV-cache management, continual learning, and reasoning under memory constraints
Inference Optimization
Efficient attention, KV-cache compression, sparse & low-rank methods for faster, leaner LLM serving at scale
Synthetic Data Curation
Generate-validate pipelines, quality filtering, model-collapse prevention, and safety-oriented data curation for LLM training
My Vision
Natural selection shapes the newborn brain's wiring and topology
Architecture & pretraining shape the model's initial weights
Guided learning forms task-based memory and skills
SFT builds task-specific skills through curated instruction
Sleep consolidates memory — replaying, pruning, strengthening
KV management consolidates context — evicting, compressing, retaining
Short-term and long-term memory store and retrieve knowledge
KV cache as dynamic context-based weights; model weights as permanent storage
Real-world feedback refines intuition and adapts behavior
Continual learning updates both KV (context) and weights (parameters)
Build tools — books, calculators — to extend cognition
RAG & tool use augment models with external knowledge
Diverse attempts, verified by outcomes, drive evolution
Diversity + verification: generate, verify, and improve
Decompose goals into subgoals and plan multi-step actions
Chain-of-thought & agentic planning decompose complex tasks
Specialized AI will develop distinct memory profiles — just as human experts develop domain intuition. Diversity with verification is how both human societies and AI systems evolve.