BEGIN:VCALENDAR
VERSION:2.0
PRODID:-//Memento EPFL//
BEGIN:VEVENT
SUMMARY:lunch&LEARN: Challenges for AI in Multimodal STEM Assessments: a H
 uman-AI Comparison
DTSTART:20250410T121500
DTEND:20250410T130000
DTSTAMP:20260416T132741Z
UID:e018a62d2613216c131d7d83dd4024e89290c1b5bd649a7843e63d85
CATEGORIES:Conferences - Seminars
DESCRIPTION:Anna Sotnikova\nGenerative AI systems have rapidly advanced\, 
 with multimodal input capabilities enabling reasoning beyond text-based ta
 sks. In education\, these advancements could influence assessment design a
 nd question answering\, presenting both opportunities and challenges.\n\nT
 o investigate these effects\, Anna Sotnikova and her team introduced a hig
 h-quality dataset of 201 university-level STEM questions\, manually annota
 ted with features such as image type\, role\, problem complexity\, and que
 stion format.\n\nTheir study analyzed how these features affect generative
  AI performance compared to students. They assessed the GPT model family u
 sing five prompting strategies and compared results to an average of 546 s
 tudent responses per question.\n\nWhile models correctly answered on avera
 ge 58.5% of questions using majority vote aggregation\, human participants
  consistently outperformed AI on questions involving visual components.\n\
 nInterestingly\, human performance remained stable across question feature
 s but varied by subject\, whereas AI performance was susceptible to both s
 ubject matter and question features.\n\nClosing the session\, Anna will pr
 ovide actionable insights for educators\, demonstrating how question desig
 n can enhance academic integrity by leveraging features that challenge cur
 rent AI systems without increasing the cognitive burden for students.
LOCATION:ME A3 31 https://plan.epfl.ch/?room==ME%20A3%2031 https://epfl.zo
 om.us/meeting/register/mMi_T7npRXOHSRu1GfuYww
STATUS:CONFIRMED
END:VEVENT
END:VCALENDAR