Neuro-Symbolic Framework Boosts LLM Agent Performance in Uncertain Environments

Large Language Model (LLM) agents are increasingly deployed in autonomous applications, but often face challenges when operating with limited information, a condition known as partial observability. This scenario introduces difficulties in inferring belief states, aligning with task objectives, and planning effectively under uncertainty. Traditional approaches that rely on full or summarized action-observation histories can often mislead LLM agents due to redundant or irrelevant information. Inspired by human cognitive processes, a new neuro-symbolic fast-slow thinking (NeSyFS) framework has been introduced to address these challenges. NeSyFS utilizes a knowledge graph to represent the agent's belief state, providing structured context to all its modules. The framework features a "fast-thinking" module for reactive actions and a "slow-thinking" module that performs uncertainty-aware planning, leveraging a twisted sequential Monte Carlo algorithm. To mitigate task objective misalignment, NeSyFS incorporates a reflection module. This module allows the agent to reflect on its fast-thinking actions and switch to the more deliberate slow-thinking mode if reactive actions repeatedly fail. Experiments conducted on benchmarks like ALFWorld, Webshop, and ScienceWorld demonstrate that NeSyFS significantly outperforms previous methods, showcasing its advantages in complex, partially observable environments.