Search Papers | Poster Sessions | All Posters
Poster B60 in Poster Session B - Thursday, August 8, 2024, 1:30 – 3:30 pm, Johnson Ice Rink
What common error patterns can tell us about human problem solving
Caroline Ahn1 (ahncj@bu.edu), Quan Do1, Leah Bakst1, Michael Pascale1, Jingxuan Guo1, Joseph McGuire1, Michael Hasselmo1, Chantal Stern1; 1Boston University
This study examines human abstract reasoning using the Cognitive Abstraction and Reasoning Corpus (CogARC), a visuospatial task inspired by an AI competition and adapted here to assess human problem-solving strategies. We analyzed online behavioral data from 233 participants who engaged in few-shot learning to learn input-output transformation rules from limited examples and apply these to novel problems. Our human subjects (M = 78.9% accuracy) significantly outperformed competing AI programs in the task. While the performance data shows considerable subject- and task-level variability, DBSCAN clustering of first attempt solutions also reveals that on certain tasks, a substantial proportion of participants made similar errors. The findings suggest shared cognitive biases in human abstract reasoning and suggest directions for future research to explore the representational space of problem-solving.
Keywords: Abstraction Reasoning Problem Solving Decision Making