Large language models (LLMs) are impressive, but they sometimes struggle with complex reasoning and can even hallucinate facts. One promising way to improve their reasoning abilities is by connecting them to knowledge graphs (KGs), which provide structured, factual information. However, existing methods often involve complex, step-by-step interactions between the LLM and the KG, which can be inefficient and limit the LLM's global planning abilities. A new research paper introduces an innovative framework called Knowledge graph Assisted Reasoning Path Aggregation (KARPA). This approach lets LLMs leverage the entire knowledge graph for reasoning, rather than just exploring it piece by piece. KARPA works in three stages: pre-planning, matching, and reasoning. First, in pre-planning, the LLM generates initial potential relationship paths between concepts based on the question. These paths are then broken down into individual relations, and similar relations are extracted from the entire knowledge graph. This helps to ensure that all potentially relevant information is considered. Next, in the matching phase, an embedding model finds paths within the knowledge graph that are semantically similar to the LLM’s proposed paths. This acts as a bridge, connecting the LLM’s understanding of the question with the structured information in the KG. Finally, during the reasoning stage, the LLM receives these matched paths from the KG and uses them to generate the final answer. The researchers found that KARPA significantly outperforms existing methods, requiring fewer interactions with the KG while improving the accuracy of the answers. This improved efficiency stems from KARPA’s ability to let the LLM consider all possible connections within the KG upfront, mimicking how humans often plan and reason. The training-free nature of KARPA also makes it easy to adapt to different LLMs and KGs, opening doors for more robust and versatile reasoning systems. While the performance of KARPA, like other LLM-based approaches, is still influenced by the quality of the LLM itself, it represents a significant step toward making LLMs more reliable and efficient reasoners by strategically integrating external knowledge.