# Resource-Aware Neuro-Symbolic Reasoning for Local Small Language Models

arXiv: https://arxiv.org/abs/2606.27281 · HTML: https://mlsanity.com/paper/2606.27281/

## Key facts

- No citing papers in the map yet (data through 14 Oct 2026).
- Placed in Reasoning benchmarks and evaluation, one of 17 directions in LLM reasoning.
- Builds on 1 of its 20 references that are in the map.
- First posted on arXiv 25 Jun 2026.

## Placed in

- LLM reasoning / [Reasoning benchmarks and evaluation](https://mlsanity.com/lr/reasoning-benchmarks-and-evaluation/)

## Builds on (most cited first, 1 of 1)

- [SmallThinker: A Family of Efficient Large Language Models Natively Trained for Local Deployment](https://mlsanity.com/paper/2507.20984/) (2025)

## Abstract (arXiv)

Small language models can run locally on consumer hardware, but structured reasoning often pushes users toward repeated sampling or larger models. This article studies a bounded neuro-symbolic alternative for local inference: a model translates a problem into typed finite-domain rules and constraints, a symbolic layer checks traceability and consistency, and a deterministic solver performs the reasoning step. The resulting Verifiable Formalization and Repair pipeline (VFR-LLM) tests when symbolic verification can replace repeated sampling without weakening accuracy. We evaluate it through LM Studio on Apple Silicon, using Qwen3-4B-2507 as the primary model, with Phi-4-mini-reasoning and Gemma-3n-E4B as robustness checks. On 120 generated pure-precedence problems, Qwen VFR-LLM achieves 0.983 accuracy, versus 0.700 for serial self-consistency using one model call instead of five. On a 120-instance BBH-derived extended logical-deduction subset, it reaches 0.933 versus 0.283. The advantage persists against a stronger cost-aware adaptive self-consistency baseline, which lowers sampling cost but not the single-call accuracy gap. Gemma reproduces the same model-dependent boundaries and Phi is negative on typed constraints. The contribution is bounded: finite-domain logic can replace repeated local sampling for some structured tasks, saving model calls and serial latency, with stratum-dependent token savings.