LLM Reasoner and Automated Planner: A new NPC approach

TL;DR

The paper addresses the challenge of creating plausible, human-like NPCs without exhaustively hand-specifying decisions for every scenario. It introduces a hybrid architecture where a Large Language Model acts as a decision-maker (Reasoner) to select goals, while a classical automated planner (AP) generates executable action sequences to achieve those goals, using a dedicated interface to integrate world state, memories, and domain definitions. Implemented in the Rhymas simulation framework on The FireFighter Problem, the system demonstrates coherent, context-aware behavior across multiple agent roles, with performance limited by LLM latency and hardware constraints. The work highlights the potential of combining LLM-based reasoning with formal planning to deliver flexible, adaptable NPCs for formative simulations and serious games, while outlining concrete future improvements such as larger models, hybrid reasoning, and broader environment testing.

Abstract

In domains requiring intelligent agents to emulate plausible human-like behaviour, such as formative simulations, traditional techniques like behaviour trees encounter significant challenges. Large Language Models (LLMs), despite not always yielding optimal solutions, usually offer plausible and human-like responses to a given problem. In this paper, we exploit this capability and propose a novel architecture that integrates an LLM for decision-making with a classical automated planner that can generate sound plans for that decision. The combination aims to equip an agent with the ability to make decisions in various situations, even if they were not anticipated during the design phase.

Figures (7)

• Figure 1: General scheme of the system behaviour. The intelligent agent receives the information about the world state and uses an LLM to generate the goal to follow; then uses PA techniques to, with the provided world state and the generated goal, find a list of actions necessary to achieve it.
• Figure 2: General scheme of The FireFighter Problem Scenario.
• Figure 3: General architecture of the intelligent agent designed and implemented.
• Figure 4: Representation of the initial state of The Simplified FireFighter Problem. In location L1 there is a burning car and in location L2 there is an extinguisher. The firefighter is in L1 an its goal is to put out the fire using the extinguisher.
• Figure 5: Complete architecture of the Intelligent Agent designed. As we can see, the Interface receives the world state's information and uses it to update the possible goals (with the help of the General Goals module), the memories of the Reasoner and the objects and predicates of the AP Problem. If a new memory has been added, the Interface says the Reasoner to set a new goal, an it is done by calling the LLM. If the AP Problem has changed since the last iteration, the Interface says the Planner to re-plan, which uses the AP Problem (and the AP Domain defined by the Domain Definition module) to create a new plan. The interface sequentially take the actions of the generated plan and gives them to the environment.