The Challenge of Partial Observability in LLM Agents

LLM-based agents often struggle in environments where they lack full information (partial observability). Standard prompting or chain-of-thought methods frequently fail because they rely on the model's internal state to track complex, changing environments, leading to hallucinations or poor long-term planning. NeSyFS (Neuro-symbolic Fast-Slow thinking) addresses this by decoupling reactive behavior from deliberate, state-based reasoning.

The Dual-Process Architecture

NeSyFS implements a cognitive architecture inspired by human 'Fast and Slow' thinking:

• Fast Thinking (Neural): This component acts as a reactive layer, utilizing the LLM's pattern recognition capabilities to make immediate, low-latency decisions based on current observations. It is optimized for speed and handling routine tasks where deep deliberation is unnecessary.
• Slow Thinking (Symbolic): When the agent encounters uncertainty or complex state transitions, it triggers a symbolic reasoning module. This module maintains a structured representation of the environment, allowing the agent to perform explicit state tracking, logical planning, and verification. By grounding the LLM's output in symbolic logic, the agent reduces the risk of drifting from the environment's constraints.

Improving Reliability through Neuro-symbolic Integration

By combining these two modes, NeSyFS allows agents to maintain a 'world model' that is updated via symbolic rules while leveraging the linguistic flexibility of neural models. This hybrid approach ensures that the agent remains grounded in the environment's reality, even when observations are incomplete or noisy. The symbolic layer acts as a constraint mechanism, preventing the neural layer from making invalid moves or losing track of critical state variables, which is a common failure mode in pure LLM-based agent implementations.