Multi-Agent Specialist Debate for Abstract Visual Concept Learning

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

Inducing a single abstract rule that separates two sets of images probes structured visual reasoning that goes well beyond describing what is in a picture, and current vision-language models are brittle at it. This work studies whether a multi-agent architecture, in which specialists examine an image along different visual dimensions and coordinate through structured debate, can reliably surface the underlying rule where a single model fails.

References

• Barrett, Hill, Santoro, Morcos & Lillicrap (2018) · Measuring abstract reasoning in neural networks (ICML)
• Chollet (2019) · On the Measure of Intelligence