Figures and short summaries from my papers, ordered by year. Select a figure to see its source.
20 paper highlights · newest first
2026 · Journal of Autism and Developmental Disorders
PaiCoach combines AI-generated feedback with human review to support parent-led communication interventions. A four-family single-case study found improvements in parent implementation and child responsiveness, while showing the continued need for expert oversight.
AIED 2026 · Companion proceedings, pp. 498–504
An LLM-powered humanoid robot combines speech, vision, and gesture during shared reading about AI. Comparing three dialogue policies on real hardware identifies how structured interaction and bounded language-model support can improve pedagogical coherence.
AIED 2026, pp. 539–554
SciEval benchmarks how well language models evaluate K-12 science instructional materials against an educator-designed rubric. It contains 273 lessons annotated across 13 criteria, with scores and supporting rationales, and tests whether domain-specific fine-tuning improves evaluation.
LAK 2026 Companion Proceedings, pp. 246–248
LLM4Art uses expert-informed prompts and rubrics to examine visual analysis, historical context, and cultural significance. A 50-artwork benchmark compares five language models, and a web application makes the interpretation workflow available to users.
Computing Conference 2026 · Intelligent Computing, pp. 481–496
Two science education experts review evaluations from GPT, Claude, and Gemini across 12 curriculum units. The study distinguishes agreement on numerical scores from agreement on reasoning and suggestions, revealing where model judgments depart from educational expertise.
AIED 2025, pp. 285–292
StoryLab helps K-2 teachers create personalized stories with illustrations, narration, and vocabulary support. Teachers specify learning goals and review generated materials, connecting children's interests with classroom literacy objectives.
AIED 2025, pp. 48–62
ASD-HI provides a multimodal benchmark for detecting parents' communication strategies and assessing implementation fidelity during home interventions. A baseline pipeline combines speech recognition, video analysis, and language-model prompting to support automated assessment.
2025 · Computers and Education: Artificial Intelligence, article 100484
Storio pairs an interactive e-book with Mia, a bilingual conversational reading partner. A design-probe study with 17 children, plus parents and educators, examines adaptive feedback and identifies design considerations for children learning English as a foreign language.
2025 · British Journal of Educational Technology, 56(5), 1784–1813
A randomized trial with 67 children compares AI-guided and parent-led e-book reading. Children in the AI-guided group showed stronger reading comprehension, while vocabulary and story-retelling benefits were comparable; the study also examines differences in engagement.
Workshop on Epistemics and Decision-Making in AI-Supported Education · AIED 2025
This study tests how the number and quality of prompt examples affect LLM assessment of science essays. Three-example GPT-4o prompts outperformed customized baseline tools for identifying rubric-based main ideas, underscoring the importance of demonstration design.
2024 · arXiv:2410.24029
Jointly trains an answer classifier and a policy for deciding when to defer to a human. Experiments on four short-answer classification tasks show that joint training improves selective prediction and the performance of both components.
Findings of EMNLP 2023, pp. 3879–3891
AsRRN models relationships among question contexts, prompts, reference answers, and student responses. Developed with 6,532 statistics responses, it combines recurrent message passing and contrastive learning to assess answers and group them for feedback.
AIED 2023, pp. 414–425
Learns when an automatic short-answer grader should defer to a human using a logistic-regression policy. The study examines how deferral choices balance assessment accuracy with the amount of human grading required.
2023 · European Journal of Applied Sciences, 11(2), 626–644
Compares classroom grades with reliable post-hoc assessments of college physics lab reports. The findings motivate paired rubrics: one to communicate expectations to students and a finer-grained version to make grading more consistent.
2022 · 11th International Conference on Teaching Statistics (ICOTS 11)
Studies AI-assisted feedback for six short-answer statistics tasks completed by 1,935 students. Human-rating reliability and exploratory response clustering provide a foundation for formative assessment in large-enrollment classes.
EMNLP 2021, pp. 6030–6040
SFRN learns relationships among a question, reference answers, and a student's answer for automatic short-answer grading. Semantic feature-wise transformations combine those relations, while translation-based data augmentation addresses limited and imbalanced training data.
AAAI 2019, 33(1), 9973–9974
MC-MDP uses reinforcement learning to select question-relevant sentences before extracting an answer. Experiments on MS MARCO show how selecting main content improves downstream reading-comprehension models.
2018 · China Conference on Information Retrieval (CCIR), pp. 93–104
A hierarchical reading-comprehension pipeline first selects passages, then question-relevant sentences, and finally an answer span. The method combines attention-based encoding with a Markov decision process and is evaluated on DuReader search data.
CIKM 2016, pp. 2437–2440
A visual dataflow system lets users build, run, monitor, and reuse machine-learning workflows as directed acyclic graphs. It connects a drag-and-drop studio with algorithm libraries and Hadoop/Spark infrastructure, including both standalone and distributed operations.
NLPCC 2014, pp. 168–180
A semi-supervised language model incorporates sentiment labels into word-vector learning. Evaluated on English and Chinese product reviews, the approach studies how semantic and sentiment information can jointly support review classification.
No papers match. Try another search or choose all years.
The full publication list also includes posters and dataset records.
I continue to work with Dr. Jinjun Xiong and serve as AI Team Lead at the following University at Buffalo centers:
Explore related papers, datasets, and systems grouped by their shared research questions.
I plan to develop a theoretical framework for translating educational goals into computational tasks, informed by interdisciplinary collaboration, and an agent-based system to support task formulation and evaluation design. I also plan to investigate responsible AI use in education, including potential risks of emotional overreliance and declines in students’ independent reasoning.
