🤖 AI

NeSyFS: A Neuro-symbolic Fast-Slow Thinking Framework for LLM Agent under Partial Observability

This paper proposes NeSyFS, a neuro-symbolic fast-slow thinking framework that utilizes knowledge graphs for belief state representation and integrates reactive actions with uncertainty-aware planning and reflection to enhance LLM agent performance under partial observability.

Original authors: Duo Xu, Faramarz Fekri

Published 2026-08-03

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Duo Xu, Faramarz Fekri

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are playing a video game where the screen is covered in fog. You can only see the few feet directly in front of your character, but you need to find a hidden treasure chest in a massive, complex castle. You can't see the whole map, and you don't know exactly where you are relative to the goal. This is the world of "partial observability" for Artificial Intelligence. In this corner of computer science, researchers are teaching Large Language Models (LLMs)—the super-smart text generators that power chatbots—to act as autonomous agents. These agents must make decisions, use tools, and solve problems, but they often have to do so without seeing the full picture. The big challenge is that when an AI can't see the whole state of the world, it gets confused. It might forget what it just did, lose track of its goal, or make a plan based on a guess that turns out to be wrong. Scientists want to build agents that can navigate this foggy world as well as a human would, using a mix of quick instincts and careful, slow thinking.

Enter a new framework called NeSyFS, which stands for Neuro-symbolic Fast-Slow thinking. Think of this as a new operating system for an AI agent designed specifically to handle that foggy video game scenario. The researchers, Duo Xu and Faramarz Fekri from Georgia Tech, realized that previous methods were like trying to remember a whole movie by reading a long, messy summary of every single line of dialogue. It's too much information, and the important details get lost in the noise. Instead, NeSyFS gives the AI a "living memory book" called a Knowledge Graph. This isn't just a list of notes; it's a structured web of facts, like a spiderweb connecting "the fridge is open" to "there is a cup inside."

The framework splits the AI's brain into two modes, inspired by how humans think. The "Fast-Thinking" mode is the agent's reflex. It glances at its memory book, grabs the most relevant facts (like "I need a cup, and I just saw one in the fridge"), and makes a quick move. But sometimes, the fog is too thick, and the quick guess is wrong. That's when the "Slow-Thinking" mode kicks in. This is the agent's deliberate planner. It doesn't just guess; it runs a simulation in its head, imagining different future paths like a chess player thinking three moves ahead. It uses a special math trick (borrowed from a method called Twisted Sequential Monte Carlo) to test out many possible futures at once, keeping the ones that look promising and discarding the ones that lead to dead ends.

To make sure the agent doesn't go off the rails, there's a third part: a "Reflection Module." Imagine a coach standing on the sidelines. Every time the agent makes a quick move, the coach checks the memory book to see if that move actually helps reach the goal. If the coach says, "No, that doesn't make sense," the agent has to rethink. If the agent keeps making the same mistake, the coach calls in the "Slow-Thinking" planner to figure out a better strategy. The paper suggests that this combination—structured memory, quick reflexes, careful planning, and a critical coach—helps the AI perform much better than previous methods. In tests on three different game-like environments (a virtual house, an online store, and a science lab), the NeSyFS agent significantly outperformed other top methods, solving tasks more often and making fewer errors. The researchers found that by using their structured memory book instead of just reading long history logs, the agent could keep its goals straight and navigate the fog much more effectively.

Technical Summary: NeSyFS – A Neuro-Symbolic Fast-Slow Thinking Framework for LLM Agents under Partial Observability

Problem Statement

Large Language Models (LLMs) are increasingly deployed as autonomous agents in complex environments such as self-reflection, retrieval-augmented generation, and scientific discovery. However, a fundamental challenge arises from partial observability: unlike supervised settings with complete input contexts, agentic environments typically reveal only local and immediate observations, while the true underlying state remains latent.

This partial observability introduces three critical challenges for LLM agents:

Belief State Inference: Agents must infer a belief over the hidden state from incomplete and potentially noisy interaction histories, which are often too long to compress effectively within limited context windows.
Task Objective Misalignment: Uncertainty and missing state information can cause agents to produce actions that deviate from the global task objective.
Planning Under Uncertainty: The inability to access the true state complicates the prediction of future situations during planning.

Prior approaches typically condition actions on full or summarized action–observation histories. These methods often rely on unstructured memory, where redundant and irrelevant information can mislead decision-making.

Methodology: The NeSyFS Framework

The authors propose NeSyFS, a novel Neuro-Symbolic Fast–Slow thinking framework designed to address these challenges in a unified manner. The framework integrates a structured memory system with a dual-process cognitive architecture inspired by human cognition (Kahneman 2011).

1. Neuro-Symbolic Memory (Knowledge Graph)

Instead of using unstructured interaction histories, NeSyFS employs a Knowledge Graph (KG) to represent the latent state of the environment.

Representation: The environment state is encoded as structured triplets (e.g., (entity, relationship, entity), (entity, attribute, value), (entity, past_tense_verb, entity)).
Dynamic Update: As the agent interacts with the environment, new observations and actions are transformed into triplets and incorporated into the KG. Outdated triplets are pruned to ensure a consistent and dynamically updated representation.
Retrieval: Every module within NeSyFS retrieves task-relevant triplets from the KG to serve as structured context, approximating the underlying state of the Partially Observable Markov Decision Process (POMDP).

2. Fast-Thinking Module (Reactive Action)

The fast-thinking module generates reactive actions efficiently.

Process: It retrieves relevant triplets from the current KG based on the task description and current observation.
Decision: Using Chain-of-Thought (CoT) reasoning, the LLM is prompted to select the next action based on this structured context.
Role: This module handles routine, locally plausible actions without the computational overhead of deep planning.

3. Reflection Module (Alignment Check)

To mitigate task objective misalignment, a reflection module evaluates the consistency of fast-thinking actions with the task goal.

Mechanism: Unlike self-reflection methods that introspect on behavior, this module adopts a third-person perspective. It uses KG-retrieved triplets as context to verify if a candidate action aligns with the task objective.
Self-Consistency: To reduce LLM randomness, the module draws multiple samples (e.g., $n=5$ ) and only approves an action if a high threshold (e.g., 80%) of samples agree.
Trigger: If an action is approved ("Yes"), it is executed. If rejected ("No"), the reflection thought is sent back to the fast-thinking module to generate a new action. If rejection occurs repeatedly (e.g., $K=3$ times), the system escalates to the slow-thinking module.

4. Slow-Thinking Module (Uncertainty-Aware Planning)

When fast-thinking is unreliable, the agent transitions to deliberative planning using a novel Neuro-Symbolic Twisted Sequential Monte Carlo (TSMC) algorithm.

Particle Representation: The algorithm maintains a set of weighted particles, where each particle represents a possible future trajectory consisting of predicted observations, latent states (KGs), and actions.
Propagation: For each particle, the LLM predicts the next action, the resulting observation, and updates the latent state (KG) accordingly.
Weight Update: Particle weights are updated based on a binary LLM evaluation of whether the state transition makes progress toward the task. The concise KG representation helps reduce hallucinations in this evaluation.
Resampling: Particles showing poor task progress are discarded, while promising ones are preserved. This mechanism makes the planning robust to temporary LLM evaluation errors.
Execution: After planning over a horizon $H$ , a particle is sampled, and its first action is executed in the actual environment.

Key Contributions

Unified Framework: A neuro-symbolic fast-slow thinking framework that addresses latent state inference, task misalignment, and planning under uncertainty simultaneously.
TSMC-Style Planning: The introduction of a TSMC-style planning algorithm within the slow-thinking module, specifically adapted for LLM agents using KG-retrieved triplets to handle uncertainty under partial observability.
KG-Augmented Reflection: A reflection module that utilizes KG-retrieved triplets to enable efficient, stepwise reflection (as opposed to trajectory-level reflection), significantly improving alignment with task objectives.
Structured Context: Demonstration that using KG-retrieved triplets as context outperforms full-history or summarized-belief approaches in decision-making, reflection, and world modeling.

Experimental Results

The framework was evaluated on three representative benchmarks: ALFWorld (embodied household tasks), WebShop (online shopping decision-making), and ScienceWorld (procedural and scientific reasoning).

Overall Performance: NeSyFS consistently and significantly outperformed state-of-the-art baselines, including ReAct, Reflexion, ABBEL, RAFA, and SwiftSage.
- On GPT-5, NeSyFS achieved an average success rate of 93.6% on ALFWorld, 63.3% on WebShop, and 69.4% on ScienceWorld, surpassing the best baseline (SwiftSage) by a significant margin.
- Similar improvements were observed with GPT-5-mini and Llama-3.3-70B.
Reflection Efficacy: Experiments isolating the reflection module showed that using KG-provided context significantly reduced Total Detection Errors (TDE) and increased Effective Reliability (ER) compared to using full interaction history or LLM-summarized belief states.
Context Comparison: In fast-thinking scenarios without reflection, KG-retrieved triplets yielded higher success rates than full history or belief summaries, validating the superiority of structured symbolic memory over unstructured text.

Significance and Claims

The paper claims that NeSyFS represents a significant advancement in LLM agent architecture for partial observability. By moving away from unstructured, redundant interaction histories toward a structured, symbolic Knowledge Graph, the framework provides a more robust and efficient representation of the latent state.

The authors highlight that they are the first to develop a TSMC-style planning method for LLM agents, leveraging the resampling mechanism to balance exploration and exploitation while remaining robust to LLM evaluation errors. Furthermore, the integration of a KG-augmented reflection module addresses the critical issue of task misalignment more effectively than conventional trajectory-wise reflection. The results suggest that the synergy between structured memory (KG), reactive fast-thinking, and deliberative slow-thinking is essential for high-performance autonomous agents in complex, partially observable environments.