Recursive Micro-Networks for Efficient Reasoning

Consider deploying a micro-network for ARC-AGI, a benchmark involving abstract visual reasoning.

Architecture Choices#

• Reasoning core: small vision transformer with shared weights across iterations.
• Verifier: lightweight classifier scoring adherence to observed transformations.
• Controller: decides when patterns stabilize or when to explore alternative hypotheses.

Workflow#

1. Convert puzzle grids into symbolic tokens describing colors, shapes, adjacency.
2. Iteratively propose transformation rules; verifier checks compatibility across examples.
3. Upon convergence, apply rule to test grid and return result with explanation.
4. If conflicting rules arise, controller backtracks using saved state snapshots.

Performance#

• Achieves competitive accuracy by focusing compute on ambiguous cases.
• Provides human-readable reasoning traces (“Detected color flip followed by rotation”).
• Maintains low latency suitable for interactive puzzle solving.