Neurosymbolic Routing for Reliable Reasoning on Resource-Constrained Edge Devices

Summary

The article proposes a neurosymbolic router that learns a deterministic finite automaton (DFA) using L* grammatical inference to classify edge device queries and route structured tasks (arithmetic, logic) to exact symbolic solvers. A small language model is used only for truly open-ended problems. On a Raspberry Pi 4B, the router achieves 98.3% overall accuracy and an 8.8x speedup compared to the Program-of-Thought baseline on 100 held-out requests.

Why it matters

As privacy, latency, and energy budgets tighten, edge and on-device inference is becoming an increasingly critical constraint. However, small language models remain fragile on tasks that are actually deterministic. This approach can enable reliable reasoning on hardware that cannot afford large models by delegating structured queries to exact engines. This reduces both energy costs and hallucination risks in offline environments.

Highlights

• L* grammatical inference automatically learns the routing DFA. In this process, the SLM itself is used as the membership oracle; rules are not hand-coded.
• Formatted queries are answered in 1–11 ms without touching the language model. This results in a 2.8% energy saving compared to Program-of-Thought.
• 98.3% overall accuracy on held-out sets from GSM8K, RuleTaker, and DeepMind Mathematics; 72.0% for PoT, and 58.7% for tool-calling agents

Critical take

The evaluation relies on only 100 unseen requests from three relatively well-structured benchmarks. Therefore, the claim of 100% routing accuracy is fragile and may lose validity on more scattered, out-of-distribution real-world queries. Additionally, the 93.3% word question accuracy still indicates a non-negligible error rate. The baselines (a single PoT implementation, a single tool-calling agent) are narrow in scope. A more aggressively prompted or fine-tuned SLM could narrow the gap. This suggests that the speedup in the headline may be partly a result of baseline selection.