Knowledge Graphs Boost Small Language Model Reasoning Capabilities

Large Language Models (LLMs) excel at reasoning but are expensive to deploy. Small Language Models (SLMs) offer a more sustainable alternative but struggle with complex, multi-hop logical reasoning. This research explores a neuro-symbolic agentic framework designed to improve SLM reasoning by integrating knowledge graph grounding. The framework transforms SLMs like Gemma and Llama into minimalist agents that use two specific tool calls: one for extracting symbolic triplets (facts) and another for obtaining expert reasoning hints via a Relational Graph Convolutional Network (RGCN). Evaluations on a kinship benchmark show that RGCN-derived hints can significantly boost performance compared to story-only baselines. However, the study also identifies critical challenges. The system's performance is constrained by errors in the initial fact extraction process, which can compound in multi-hop reasoning chains. Additionally, some SLM architectures exhibit a "distraction effect," where noisy, self-generated facts degrade performance even when expert hints are available. This work provides insights into the difficulties of symbolic grounding in low-resource agentic systems and suggests a roadmap for iterative verification in neuro-symbolic pipelines.