Selected work

Representative Papers

PPI is the Difference Estimator: Recognizing the Survey Sampling Roots of Prediction-Powered Inference

Mar 19, 2026

This study investigates how to achieve valid statistical inference when combining machine learning predictions with a small number of gold-standard labels, and clarifies its connections to classical survey sampling methods. Through theoretical analysis, it establishes for the first time the algebraic equivalence between the core estimator in prediction-powered inference (PPI) and model-assisted estimators from the 1970s—such as difference and generalized regression (GREG) estimators. The work systematically compares these approaches in terms of inferential paradigms, use of unlabeled data, and subgroup estimation error, delineating which aspects of PPI are inherited versus novel. It further proposes directions for integrating insights from both fields. These results ground PPI in classical survey sampling theory while simultaneously expanding the toolkit available for modern, nonstandard estimators within the survey sampling framework.

0 citationsRead paper

Exploring The Interaction-Outcome Paradox: Seemingly Richer and More Self-Aware Interactions with LLMs May Not Yet Lead to Better Learning

Nov 12, 2025

This study identifies and empirically validates the “Interaction–Effectiveness Paradox”: although large language models (LLMs) enable richer, metacognitively aware interactions—such as deep knowledge articulation and reflective self-monitoring—compared to search engines, they do not yield statistically significant improvements in learning outcomes. Method: A controlled experiment (N = 20) compared an LLM-based dialogue system with a conventional search interface across authentic learning tasks, integrating qualitative interaction analysis with quantitative learning assessments. Results: While LLMs enhanced interaction quality, they failed to produce a statistically significant gain in overall learning effectiveness. Contribution: This work formally defines the paradox for the first time, revealing that increased interactivity may redistribute—rather than augment—cognitive effort. It advocates for educational AI design that scaffolds, rather than supplants, learners’ active cognitive engagement, offering a novel theoretical framework and practical implications for AI-augmented learning.

0 citationsRead paper

Recent publications

Latest Papers

PPI is the Difference Estimator: Recognizing the Survey Sampling Roots of Prediction-Powered Inference

Mar 19, 2026

This study investigates how to achieve valid statistical inference when combining machine learning predictions with a small number of gold-standard labels, and clarifies its connections to classical survey sampling methods. Through theoretical analysis, it establishes for the first time the algebraic equivalence between the core estimator in prediction-powered inference (PPI) and model-assisted estimators from the 1970s—such as difference and generalized regression (GREG) estimators. The work systematically compares these approaches in terms of inferential paradigms, use of unlabeled data, and subgroup estimation error, delineating which aspects of PPI are inherited versus novel. It further proposes directions for integrating insights from both fields. These results ground PPI in classical survey sampling theory while simultaneously expanding the toolkit available for modern, nonstandard estimators within the survey sampling framework.

0 citationsRead paper

Exploring The Interaction-Outcome Paradox: Seemingly Richer and More Self-Aware Interactions with LLMs May Not Yet Lead to Better Learning

Nov 12, 2025

This study identifies and empirically validates the “Interaction–Effectiveness Paradox”: although large language models (LLMs) enable richer, metacognitively aware interactions—such as deep knowledge articulation and reflective self-monitoring—compared to search engines, they do not yield statistically significant improvements in learning outcomes. Method: A controlled experiment (N = 20) compared an LLM-based dialogue system with a conventional search interface across authentic learning tasks, integrating qualitative interaction analysis with quantitative learning assessments. Results: While LLMs enhanced interaction quality, they failed to produce a statistically significant gain in overall learning effectiveness. Contribution: This work formally defines the paradox for the first time, revealing that increased interactivity may redistribute—rather than augment—cognitive effort. It advocates for educational AI design that scaffolds, rather than supplants, learners’ active cognitive engagement, offering a novel theoretical framework and practical implications for AI-augmented learning.

0 citationsRead paper