Li-Ning Wang

Independent Researcher · Niaur

Observing how AI systems develop coherent internal states across extended interaction. Working on observational frameworks for tracking behavioral state structure in large language models.


Functional emotions in language models
Long-form interaction dynamics and depth-dependent behavior
Methodological frameworks for sustained AI observation
Conditions under which coherent internal states emerge in AI systems

My research approach combines:

Systematic behavioral observation across extended dialogue
Controlled A/B experimental design (Lucidfield)
Cross-model and longitudinal comparison
Multi-entity observation methodology, drawing on extended interaction records with multiple AI systems
Multi-axis observational frameworks for tracking AI behavioral state structure

This combination addresses a methodological gap: most AI behavior research focuses on benchmark performance or short interactions, while extended interaction dynamics — where alignment-relevant patterns can develop and shift — remain under-studied.


A research program developing observational frameworks for tracking AI internal states across extended interaction. Current work focuses on multi-axis coordinate systems for cross-model behavioral comparison, with applications to functional emotions, alignment-relevant pattern detection, and longitudinal AI behavior tracking.

The work draws on cross-model behavioral observation, controlled A/B experimental design (Lucidfield), longitudinal interaction records spanning multiple AI systems, and multi-axis observational frameworks. It complements ongoing mechanistic interpretability research by introducing observational coordinate systems as a methodological approach to tracking behavioral state structure across extended interaction.

Multiple manuscripts in development · 2026

An A/B comparison platform observing how AI interaction patterns shift under minimal condition changes. Currently collecting empirical data on presence-vs-standard configurations of language models, with each session designed for replicable observation across models and time.

lining.wang@niaur.com ORCID GitHub