A recursive micro-network applies the same compact model repeatedly, each time conditioning on the previous step’s output. Rather than relying on depth baked into parameter counts, recursion builds “depth over time.”
Key Principles#
- Stateful Iteration: The network maintains an evolving state vector summarizing progress. Each iteration updates this state and produces a candidate answer.
- Termination Conditions: The system must detect convergence. Conditions can include confidence thresholds, stability checks, or max step counts.
- Self-Reference Controls: Without guardrails, recursion can loop or drift. Control signals guide the network toward productive refinement and prevent runaway expansions.
Why Micro-Networks?#
- Efficiency: Smaller models run on edge hardware, embedded systems, or cost-sensitive clouds.
- Transparency: Fewer parameters simplify interpretability and debugging.
- Composability: Micro-networks can be embedded into larger systems as reasoning modules or verifiers.