2026 AI-ALOE Mini Retreat

September 14, 2026

The 2026 AI-ALOE Mini Retreat was held virtually on September 11, with about 50 participants, including research fellows, teaching fellows, and the external advisory board. Discussions centered on evaluation and impact, with a particular focus on learning proficiency.

Our SMART team (Dr. Min Kyu Kim, Jinho Kim, Yoojin Bae) participated as one of the seven technology teams of AI-ALOE. During the first half of the session, Jinho Kim shared the approaches our team has used to measure impact, experimental design, and learning analytics. In the second half, Dr. Min Kyu Kim and Yoojin Bae focused on learning proficiency - how our team understands it and how we could measure it going forward through experimental design and learning analytics.

Read more about it here: https://lnkd.in/p/e8NW_QMf 

New Manuscript in Assessing Writing

September 12, 2026

We have a new manuscript published in Asessing Writing by Dr. Min Kyu Kim, Hyunkyu Han, Seora Kim, and Dr. Mohamed Shameer Abdeen.

Kim, M. K., Han, H., Kim, S., & Abdeen, M. S. (2026). AI-scaffolded summary writing for pre-class learning in an undergraduate physics course. Assessing Writing, 70, 101093. https://doi.org/10.1016/j.asw.2026.101093

Abstract:

Summary writing is a common write-to-learn strategy in undergraduate STEM education, particularly for pre-class learning. Yet producing summaries that demonstrate deep comprehension is demanding and often requires instructional support. In response, Automated Summary Evaluation (ASE) tools have been developed to provide formative assessment and feedback on student summaries. This study examines the effectiveness of a generative AI-powered ASE tool designed to scaffold student engagement in pre-class summarization. We investigated whether AI support enhances editing behaviors and promotes concept learning while interacting with learner background characteristics. We analyzed 1081 revision attempts across seven topics over seven weeks from 49 undergraduates in an Introductory Physics course at a large public university. A longitudinal analytic approach using Linear Mixed-Effects Models was employed to address the research questions. Findings indicate that AI-powered formative feedback fostered revision behaviors associated with higher concept learning scores. Effective revisions occurred when students added concepts in response to AI feedback while avoiding careless deletions and surface-level sentence changes. Engagement and performance varied more by assignment than by tool proficiency, with AI scaffolds especially beneficial for students historically underperforming in STEM. These results underscore the importance of personalized feedback strategies that promote targeted revision across diverse learners.

Hyunkyu Han and Jinho Kim Elected to the 2026-2027 Board of ILSSA

September 1, 2026

Hyunkyu Han and Jinho Kim has been elected to the 2026–2027 board of the International Learning Sciences Student Association (ILSSA) within the International Society of the Learning Sciences (ISLS)!

Hyunkyu Han has been elected Asia-Pacific Regional Rep, and Jinho Kim as Co-Chair. Hyunkyu will serve a one-year term, and Jinho a two-year term.

In addition, both will represent ILSSA on other ISLS committees: Jinho Kim on the Annual Meeting Committee, and Hyunkyu on the Membership Committee.

Find out more about the ISLS committees here: https://www.isls.org/members/committees/

Jinho Kim's 2026 AI4Ed-Funded Summer Graduate Fellowship & Summit

August 20, 2026

Our graduate associate, Jinho Kim, was awarded the 2026 AI4Ed-Funded Summer Graduate Fellowship and participated in the Summit following it, held August 18-19, 2026, at the Illini Center in Chicago, IL, where she presented her team's work to advisors and peers. As part of the fellowship, Jinho Kim collaborated with two others to explore multimodal models' science diagram understanding on a benchmark of 100 diagram pairs.

The AIVO AI Institute Summer Graduate Fellowship provides a 12-week program through funding from Google.org, with fellows from five AI Institutes focused on education contributing 40 hours per week.

Edu-MLLMs: Do Multimodal Foundation Models Understand or Just Trust the Diagram?

Heather Broome, Jinho Kim, Alexander Stone

Multimodal large language models (MLLMs) are increasingly proposed as automated quality control for educational materials, yet it remains unclear whether they evaluate science diagrams with any awareness of what a diagram is meant to teach and to whom. We investigate this question with a benchmark of 100 diagram pairs drawn from the AI2D-RST corpus, in which each authentic diagram is hand-edited to introduce exactly one error, yielding pairs of correct and incorrect versions. Models (Claude Sonnet 4.6, GPT-5, Gemini 2.5 Pro, Qwen2.5-VL at 7B and 72B scales, Gemma 8B, and Ministral 8B) judge every diagram under five presentation conditions that systematically vary whether the model receives the image, a statement of the intended audience and concept, and a caption that agrees or conflicts with the image, producing structured verdicts with evidence attribution and confidence for 7,000 trials in total. This design lets us measure whether learning context improves error detection and reduces false alarms on correct diagrams, whether detection depends on an error types, and which modality a model trusts when text and image contradict each other. The addition of learning context alone did not improve the models’ judgments, and when caption and image disagreed, every frontier model trusted the text over the diagram, suggesting that current MLLMs read about diagrams more than they understand them. 
 

Check out more here:

• AIVO Grad Fellow Summit 2026 
• https://lnkd.in/p/ezkEH4eK 
• https://aiinstitutes.org/august-2026-aivo-newsletter/