Keynotes | K&Ts | GACs | Talks | Posters | Search

Poster A in Poster Session A: Tuesday, August 4, 9:30 – 11:15 am, Kimmel Center, Shorin & Rosenthal Rooms

Adaptive Online Learning of Structured Abstractions in Visual Problem-Solving

Ivan Zareski¹, Luisa El Amouri¹, Pinzhe Zhao², Leonardo Hernandez Cano³, Emanuele Sansone³, Bonan Zhao², Marta Kryven¹; ¹Dalhousie University, ²University of Edinburgh, ³Massachusetts Institute of Technology

Presenter: Ivan Zareski

People solve complex problems by breaking them up into simpler structured sub-tasks, and can do so during online interaction. Which computational principles explain human sub-task discovery? We study this question in a visual Pattern Builder Task (PBT), where participants reconstruct a sequence of complex patterns from a small set of geometric primitives and transformation operations. PBT allows saving any patterns as helpers, which supports reuse of structural abstractions across sequential tasks. We find that human solutions often deviate from the shortest programs, measured in terms of raw primitives. At the same time, people reused helpers to minimize the length of compositional solutions, while growing their helper library over time. Computational modeling formalizing pattern discovery as program synthesis further revealed that human library learning serves to reduce the size of the underlying search space. Comparison between library learning algorithms based on greedy compression and based on LLM-driven program synthesis shows that only the latter replicates key behavioral signatures. This result suggests that human library learning is at least partly driven by inductive biases, as modeled by LLM-based program synthesis.

Topic Area: Memory, Learning & Knowledge Structures