Contribution:

I innovated my Bridged Clustering Algorithm to align unsupervised sources for cross-modal representation learning, developing method-agnostic, interpretable, and bidirectional models with theoretical and empirical strengths in large-scale NLP experiments.

Contribution:

I independently invented Bridged Clustering, one of the first semi-supervised learning algorithms that jointly exploits unsupervised inputs and outputs, realigning disjoint datasets towards new supervised objectives, across any pair of under-supervised modalities.

In collaboration with Stanford Biology, I experimented with my algorithm in scientific research, jointly modeling disjoint biological data collected by Dr. Wu's team.

Contribution:

I invented an Analogy Tutor system where LLM tutors teach human learners abstract concepts through cross-domain analogies, with an end-to-end concept anonymization framework and LLM-based student simulation.

Contribution:

I co-developed Social-Contract AI, a framework aligning LLM assistants with implicit societal norms such as sycophancy and altruism. I led the development of a full experimental pipeline: Constitutional AI and Verbal RL, prompt design, human and LLM-based evaluations, and large-scale data collection.

Contribution:

I co-led this study of AI technology's social integration process into families and impacts on children’s psycho-social formation, by studying how AI co-create symbolically meaningful interactions with children. This work is one of the first formulations of symbolic interactionism in Human-centered AI.

Contribution: 

Integrating in-context learning, fine-tuning with distillation, and Verbal Reinforcement Learning for semantic grouping tasks, achieving near-human-level performance and a significant improvement of 38.23% compared to baseline clustering methods.