LLM-ARK: Knowledge Graph Reasoning Using Large Language Models via Deep Reinforcement Learning

With the evolution of pre-training methods, large language models (LLMs) have exhibited exemplary reasoning capabilities via prompt engineering. However, the absence of Knowledge Graph (KG) environment awareness and the challenge of engineering viable optimization mechanisms for intermediary reasoning processes, constrict the performance of LLMs on KG reasoning tasks compared to smaller models. We introduce LLM-ARK, a LLM grounded KG reasoning agent designed to deliver precise and adaptable predictions on KG paths. LLM-ARK utilizes Full Textual Environment (FTE) prompts to assimilate state information for each step-sized intelligence. Leveraging LLMs to richly encode and represent various types of inputs and integrate the knowledge graph further with path environment data, before making the final decision. Reframing the Knowledge Graph (KG) multi-hop inference problem as a sequential decision-making issue, we optimize our model using the Proximal Policy Optimization (PPO) online policy gradient reinforcement learning algorithm which allows the model to learn from a vast array of reward signals across diverse tasks and environments. We evaluate state-of-the-art LLM(GPT-4) and our method which using open-source models of varying sizes on OpenDialKG dataset. Our experiment shows that LLaMA7B-ARK provides excellent results with a performance rate of 48.75% for the target@1 evaluation metric, far exceeding the current state-of-the-art model by 17.64 percentage points. Meanwhile, GPT-4 accomplished a score of only 14.91%, further highlighting the efficacy and complexity of our methodology. Our code is available on GitHub for further access.