Abstract:
Large Language Models (LLMs) have rapidly advanced their ability to answer questions and perform complex reasoning tasks. However, they often generate factual inaccuracies and hallucinations because they lack access or have limited access to up-to-date factual knowledge. To mitigate this, researchers often augment LLMs with factual information from external sources, such as knowledge graphs (KGs). However, most existing KG-based RAG systems suffer from a key limitation: triplet retrieval from KGs is either based on simplistic distance metrics, heuristics, or tightly coupled with reasoning, making optimizing both retrieval and reasoning challenging. To mitigate these, we propose KG-Scout , a reinforcement learning (RL)-based policy network that decouples retrieval from reasoning, enabling the selection of triplets that are both semantically aligned with the query and structurally important in the KG. Our approach operates in two key stages: (1) extracting a subgraph using topic entities and computing Personalized PageRank (PPR) scores for nodes, and (2) employing our policy network to select the most valuable triplets from this set based on their learned relevance scoring. To enhance the efficiency of this process, we first perform an initial filtering of candidate triplets using cosine similarity with the query before the policy network considers them. Using the retrieved results, smaller pretrained LLMs such as LLAMA-3.1-8b outperform several complex LLM-based baselines on WebQSP and CWQ benchmarks.
