GraphMind: LLMs as Dynamic Knowledge Builders for Sequential Decision-Making

ICLR 2026 Conference Submission16193 Authors

Keywords: Large Language Models, Sequential Decision Making

Abstract: While the reasoning capabilities of large language models (LLMs) have advanced considerably due to their extensive internal knowledge, efficiently internalizing and leveraging new information in dynamic environments remains challenging. This limitation is particularly pronounced in partially observable environments, which require agents to manage long-term memory and perform effective exploration under incomplete information. To address this, we propose an LLM agent architecture that integrates a knowledge graph as a graph-based memory module to facilitate high-level action planning. The agent incrementally constructs the knowledge graph through environmental interactions and retrieves relevant information to formulate efficient plans. We evaluate our approach in complex navigating environments specifically designed to present long-horizon and partially observable challenges. Experimental results demonstrate that employing a knowledge graph as an external memory significantly enhances the success rate and efficiency of the LLM’s planning capabilities.

Primary Area: foundation or frontier models, including LLMs

Submission Number: 16193