NeSyFS: Neuro-Symbolic Fast-Slow Thinking Framework for LLM Agents
A new framework, NeSyFS, has been proposed to address the challenges of partial observability in large language model (LLM) agents. The framework, detailed in a paper on arXiv (2607.28942), combines neuro-symbolic reasoning with a fast-slow thinking paradigm inspired by human cognition. It uses a knowledge graph to represent the agent's belief state, providing triplets as context for each module. The approach aims to improve decision-making in applications such as self-reflection, retrieval-augmented generation, and scientific discovery, where agents must act based on limited observations. The paper highlights issues like belief state inference, task objective misalignment, and planning under uncertainty, which prior approaches fail to address due to reliance on full or summarized action-observation histories containing redundant information. NeSyFS offers a unified solution to these problems.
Key facts
- The framework is named NeSyFS (Neuro-symbolic Fast-Slow Thinking).
- It is designed for LLM agents operating under partial observability.
- The paper is available on arXiv with ID 2607.28942.
- It uses a knowledge graph to represent belief states.
- The framework is inspired by human cognition's fast and slow thinking.
- Target applications include self-reflection, retrieval-augmented generation, and scientific discovery.
- It addresses challenges such as belief state inference, task objective misalignment, and planning under uncertainty.
- The approach contrasts with prior methods that use full or summarized action-observation histories.
Entities
Institutions
- arXiv