Multi-Agent Specialist Debate for Abstract Visual Concept Learning Across Bongard Problem Taxonomies

The crisis

• Visual concept learning — the ability to induce abstract rules from examples — is foundational to AI systems that must adapt to novel situations without retraining
• Current VLMs show brittle performance on structured visual reasoning; they describe images rather than induce the rule that separates them
• Multi-agent debate is underexplored for visual reasoning tasks where decomposition by visual dimension is principled, not arbitrary

About this research

Bongard problems require inducing a single abstract rule that separates two sets of images — a task that probes structured visual reasoning beyond description. This work studies whether a multi-agent architecture with dimension-specialized specialists (geometric form, spatial relationships, cardinality, position, magnitude) and a debate mechanism can reliably surface the correct concept where monolithic VLMs fail.

Key findings

• (In Progress)