Enhancing Small Language Models Reasoning through Knowledge Graph Grounding

2026-07-14

AI Summary

This work addresses the limitations of small language models in complex multi-hop logical reasoning, which often stem from insufficient symbol grounding capabilities. The authors propose a minimalist neuro-symbolic agent framework that integrates triplet-based fact extraction (extract_facts) with relational graph convolutional network (RGCN)-guided expert prompting (get_hint) to enhance the reasoning performance of compact models such as Gemma 3 and Llama 3.2. Evaluated on the CLUTRR kinship reasoning benchmark, the approach demonstrates that RGCN-derived hints can improve accuracy by 1.5–2 times. However, noise introduced during self-supervised fact extraction triggers a “distraction effect,” causing early errors to propagate and accumulate significantly across reasoning steps. These findings highlight critical challenges, including bottlenecks in symbolic extraction fidelity and the inherent fragility of sequential reasoning in current small-scale architectures.

Abstract

Although large language models (LLMs) have set benchmarks for zero-shot reasoning, their deployment remains cost-prohibitive and environmentally taxing. Small Language Models (SLMs) offer a sustainable alternative, but prone to errors, on tasks requiring complex, multi-hop logical grounding. We investigate a neuro-symbolic agentic framework to enhance the reasoning capabilities of SLMs, specifically Gemma 3 (1B, 4B) and Llama 3.2 (3B), using the CLUTRR kinship benchmark. Our approach transforms the SLM into a minimalist agent utilizing two specialized tool calls: extract_facts for symbolic triplet extraction and get_hint for expert reasoning via a Relational Graph Convolutional Network (RGCN). We evaluate these models across two configurations, both in an Oracle scenario with ground-truth triplets and a Realistic scenario relying on self-extracted knowledge. Our results reveal that while RGCN-derived hints provide a 1.5 - 2x performance gain over story-only baselines, the system is constrained by the extraction bottleneck and sequential deductive fragility, where early extraction errors compound over multi-hop chains. Furthermore, we identify a "distraction effect" in specific architectures where noisy, self-generated facts degrade performance despite the presence of expert hints. This work characterizes the challenges of symbolic grounding in low-resource agentic systems and provides a roadmap for iterative verification in neuro-symbolic agentic pipelines.

Problem

• Small Language Models
• Symbolic Grounding
• Multi-hop Reasoning
• Knowledge Graph
• Neuro-symbolic Reasoning

Innovation

• neuro-symbolic reasoning
• small language models
• knowledge graph grounding
• relational graph convolutional network
• multi-hop reasoning