XUAN LIU

I like following questions wherever they take me. People, language, numbers, technology.

Mostly, I’m just trying to understand the world a little better.

Xuan Liu Close

SPAR (Supervised Program for Alignment Research)

– Independently proposed framing AI tutoring as a partially observable decision problem, emphasizing learner-state uncertainty and evidence-seeking before intervention; co-developed a taxonomy of 14 tutoring-agent failure modes spanning over- and under-intervention.

– Designed and tested a Python-based block-coding maze environment for AI-to-teacher oversight. A REINFORCE policy outperformed tested fixed baselines while requesting teacher assistance at 17% of decision points, highlighting the limitations of simulated learner assumptions.

– Collaborated on developing and training a step-level ASK/PLAY mathematics tutoring environment using MATH and PRM800K problems; investigated how reward optimization can favor task completion over independent student learning.

– Curated 200 mathematics problems comprising 1,152 solution steps from PRM800K and designed a teacher-labeling protocol to evaluate step importance and appropriate AI assistance, preparing for future comparisons between computational labels and educators’ judgments.