A new router can tell in milliseconds whether a question actually needs a language model at all.
A preprint posted this month on arXiv (arXiv:2609.35833, not yet peer-reviewed) describes a neurosymbolic router built for small language models running on constrained edge hardware. The system classifies each incoming query and sends structured problems, arithmetic, algebra, formal logic, to deterministic symbolic solvers, reserving the language model for genuinely open-ended word problems. Rather than hand-coding that routing logic, the researchers learned it automatically with the L* grammatical inference algorithm, using the small model as an oracle during training. Tested on a Raspberry Pi 4B with no GPU, on 100 prompts from DeepMind Mathematics, GSM8K, and RuleTaker, the router hit 100% routing accuracy and 98.3% overall accuracy, against 72.0% for the strongest prior agent, Program-of-Thought, and 58.7% for a standard tool-calling agent.
The bigger point is not the accuracy jump, it is where the gains come from. Instead of squeezing more reasoning out of a small model through better prompting or bigger context windows, the router simply stops asking the model to do arithmetic it was never built to do reliably. Because formatted queries never reach the model, they get answered in 1-11 milliseconds, and the lightest configuration runs 8.8x faster and 2.8x more energy-efficient than the previous best agent approach.
It is a single, not-yet-peer-reviewed preprint tested on 100 prompts, so treat the numbers as promising rather than settled. Still, the core insight, that a lot of what looks like reasoning is really just math wearing a word problem's clothes, is a cheap trick that phone and gadget makers chasing on-device AI will likely want to steal.