Back to feed
AR
arXiv CS.AI
7/17/2026
Enhancing Small Language Models Reasoning through Knowledge Graph Grounding

Enhancing Small Language Models Reasoning through Knowledge Graph Grounding

Short summary

This study enhances small language model reasoning by grounding them in knowledge graphs via a neuro-symbolic agentic framework. Gemma 3 (1B, 4B) and Llama 3.2 (3B) are transformed into minimalist agents using extract_facts and get_hint tool calls, with an RGCN providing expert reasoning hints. RGCN hints yield 1.5-2x gains over baselines, but extraction bottlenecks and a distraction effect from noisy self-generated facts limit performance on multi-hop chains.

  • •Neuro-symbolic agentic framework grounds SLMs (Gemma 3, Llama 3.2) via knowledge graph tool calls on CLUTRR benchmark
  • •RGCN-derived hints provide 1.5-2x performance gain but extraction bottlenecks and deductive fragility limit multi-hop accuracy
  • •Identifies a distraction effect where noisy self-generated facts degrade SLM performance despite expert hints

Generated with AI, which can make mistakes.

#research-breakthrough#ai-agents#ai-tools
Read full article at arXiv CS.AI

Is this a good recommendation for you?

Comments

Failed to load comments. Please try again.

Explore more

AgentKGV: Agentic LLM-RAG Framework with Two-Stage Training for the Fact Verification of Knowledge Graphs
arXiv cs.CL
HG-RAG: Hierarchy-Guided Retrieval-Augmented Generation for Structured Knowledge Graphs
arXiv CS.AI
Grounded Iterative Language Planning: How Parameterized World Models Reduce Hallucination Propagation in LLM Agents
arXiv CS.AI
The user wants me to rewrite a news headline for a mobile feed based on the given rules.
arXiv CS.AI