A comprehensive survey published May 23 in Intelligent Computing , a Science Partner Journal maps out the role of large language models in task planning, underscoring the growing influence of artificial intelligence in complex decision-making tasks.
Traditionally reliant on expert systems and manual configuration, task planning is essential for organizing action sequences to achieve defined goals, and is now being redefined by the advanced reasoning capabilities of large language models. The survey offers a comprehensive synthesis of how these models are reshaping the field, highlighting a dual-path framework: harnessing large language models' intrinsic reasoning and integrating them with external methodologies.
Within their intrinsic capabilities, large language models are employed for step-by-step reasoning methods such as Chain of Thought and its advanced variants like Tree of Thoughts and Graph of Thoughts, enabling the decomposition of complex tasks and exploration of multiple reasoning paths. Techniques such as self-consistency and CRITIC introduce iterative feedback and external validation to improve performance, while knowledge enhancement approaches use both internal historical data and external knowledge such as Retrieval-Augmented Generation to enrich the planning process.
Beyond intrinsic reasoning, the integration of large language models with classical tools like the Planning Domain Definition Language (PDDL), reinforcement learning frameworks and interdisciplinary theories such as theory of mind, game theory, neuroscience and graph theory further expands their application. Specifically, when combined with tools like PDDL, large language models can act as planners or translators in dynamic environments.
The real-world impact of these advances is already being seen in applications such as embodied AI, game playing, and economy and society simulation. From embodied agents interacting with their physical or simulated environments to systems like ChatDev and WarAgent simulating collaborative software development and historical geopolitical dynamics, large language models are enabling machines to plan, adapt and interact in increasingly sophisticated ways.
Looking ahead, the survey identifies key challenges and opportunities, such as improving multimodal situational awareness, integrating human feedback for safety and domain expertise, enhancing real-time adaptability and developing more nuanced evaluation metrics. Aiming to serve as a foundational resource for the research community, the survey also offers a continuously updated resource hosted on GitHub ( https://github.com/ZhaiWenShuo/Survey-of-Task-Planning ).
