Sangziwang

SangZi Wang is an independent researcher focused on LLM behavioral reliability, interaction dynamics, and runtime observability. With a background in reconstructive and plastic surgery, his work bridges long-term human-AI interaction observation with practical auditing methodologies for large language models. His current research explores how conversational environments, protocol structures, and long-context interactions influence model behavior over time. Core topics include: - behavioral drift in extended dialogue, - protocol-induced response distortion, - execution confidence vs epistemic confidence (EC–EpC gap), - cross-model behavioral comparison, - runtime observability and interaction ecology. Rather than focusing on AGI speculation, his work emphasizes measurable behavioral phenomena emerging in real-world interaction settings. He has released multiple public Zenodo preprints and datasets, maintains structured GitHub repositories, and develops open behavioral audit frameworks designed for reproducible evaluation across different LLM systems. His long-term goal is to build lightweight, modular infrastructures for AI behavioral auditing, interaction reliability analysis, and runtime governance that remain accessible to independent researchers outside large institutional labs.

Personal Website

-

LinkedIn

-

Twitter / X

-

LessWrong

-

EA Forum

-

Projects

No grants recorded.