Neurosymbolic router achieves 98.3% accuracy on edge hardware by dispatching structured tasks to symbolic solvers

Researchers propose a neurosymbolic router that classifies incoming queries and dispatches them to the most efficient correct solver, sending structured arithmetic, algebra, and formal logic tasks to deterministic engines and reserving a small language model (SLM) for open-ended word problems.

The routing logic is learned as a deterministic finite automaton (DFA) using the L* grammatical inference algorithm, with the SLM serving as a membership oracle and labeled data as an equivalence oracle.

Evaluated on a Raspberry Pi 4B (8 GB RAM, no GPU) with 100 untested prompts from DeepMind Mathematics, GSM8K, and RuleTaker, the learned router achieves 100% routing accuracy and 98.3% overall accuracy with a 512-token reasoning budget.

On word problems specifically, the router attains 93.3% accuracy, outperforming the strongest agent baseline, Program-of-Thought (72.0%), and a tool-calling agent (58.7%) given the same solvers.

Because formatted queries are never passed to the model, the router answers them in 1–11 milliseconds, and in its 30-token configuration it runs 8.8 times faster and 2.8 times more energy-efficient than Program-of-Thought.