Source-linked AI summary
Understanding the planning of LLM agents: A survey
Xu Huang, Weiwen Liu, Xiaolong Chen, Xingmei Wang, Hao Wang, Defu Lian, Yasheng Wang, Ruiming Tang, Enhong Chen
TL;DR
LLM-agent planning is a growing area without a comprehensive systematic view of methods that improve planning ability. This survey organizes the literature into five directions, analyzes representative approaches, evaluates methods on four benchmarks, and discusses challenges and limitations.
Problem
Existing surveys cover related LLM capabilities and autonomous agents, but this paper identifies a need for comprehensive analysis specifically focused on LLM-based agents’ planning abilities.
Method
The survey proposes a taxonomy of five planning directions and analyzes representative works, while evaluating several methods on four benchmarks.
Results
The survey reports that representative methods improve planning performance, with performance increasing alongside expenses, and highlights promising improvements on complicated tasks such as ALFWorld and ScienceWorld.
Takeaways & Limitations
LLM-agent planning research can be viewed through task decomposition, multi-plan selection, external modules, reflection and refinement, and memory augmentation, which are interconnected directions.
Takeaways & Limitations
Reflection-based textual updates lack a guaranteed convergence proof, so continual reflection cannot currently be demonstrated to converge.
Abstract
from arXiv · showhide
As Large Language Models (LLMs) have shown significant intelligence, the progress to leverage LLMs as planning modules of autonomous agents has attracted more attention. This survey provides the first systematic view of LLM-based agents planning, covering recent works aiming to improve planning ability. We provide a taxonomy of existing works on LLM-Agent planning, which can be categorized into Task Decomposition, Plan Selection, External Module, Reflection and Memory. Comprehensive analyses are conducted for each direction, and further challenges for the field of research are discussed.
1 Introduction
Planning is a critical agent capability, but conventional symbolic and reinforcement-learning approaches face practical limitations. This survey positions LLMs as potential cognitive cores for agent planning and systematically analyzes representative methods.
- 1 Introduction: Agent planning requires understanding, reasoning, and decision-making to generate action sequences for task goals.The formulation uses the environment, goal, LLM parameters, and prompts to produce a sequence of actions.
- 1 Introduction: Symbolic methods require converting flexible natural-language problems into symbolic models, often with human-expert effort and low error tolerance.The passage states that even a few errors can cause failure.
- 1 Introduction: Reinforcement-learning methods commonly require many environment interactions, which can be impractical or costly when data collection is time-consuming or expensive.These methods may combine deep models serving as policy or reward networks.
- 1 Introduction: LLMs have demonstrated capabilities in reasoning, tool usage, planning, and instruction-following, motivating their use as agents’ cognitive cores.The survey describes this development as a paradigm shift.
- 1 Introduction: The survey systematically analyzes LLM-based agent planning, evaluates representative methods on four benchmarks, and organizes the paper around five research directions.The paper presents itself as the first comprehensive analysis of LLM-based agents from the perspective of planning abilities.
2 Taxonomy
The survey organizes LLM-agent planning into five interconnected directions and examines their motivations, representations, and limitations. These directions range from decomposing tasks and selecting plans to using external modules, reflection, and memory.
- 2 Taxonomy: The taxonomy divides LLM-agent planning into task decomposition, multi-plan selection, external module-aided planning, reflection and refinement, and memory-augmented planning.The five directions are interconnected rather than mutually exclusive and may be adopted concurrently.
- Task Decomposition: Task decomposition applies divide and conquer by splitting complex tasks into sub-tasks and sequentially planning for each one.Its process contains decomposition and sub-planning steps.
- Multi-plan Selection: Multi-plan selection generates alternative plans and uses a task-related search strategy to choose one for execution.The search component can include tree-search algorithms.
- External Planner-Aided Planning: External planner-aided planning uses an external planner to address efficiency and infeasibility while the LLM mainly formalizes tasks.The formalized information is passed to the external planner module.
- Reflection and Refinement: Reflection and refinement improves planning by having the LLM reflect on failures and refine the plan.The process iteratively produces reflection feedback and an updated plan.
- Memory-Augmented Planning: Memory-augmented planning retrieves stored commonsense, past-experience, or domain-specific information as auxiliary signals during planning.The memory module supplies information to the planning process.
3 Task Decomposition
Task decomposition addresses complex, variable environments by dividing tasks into sub-tasks and planning those sub-tasks through either decomposition-first or interleaved strategies. These approaches improve planning structure and fault tolerance but add computational overhead and can strain context length or cause long-trajectory hallucinations.
- Task Decomposition: Task decomposition breaks complex tasks into simpler sub-tasks, followed by planning for each sub-task.The process has a “decompose” step and a “sub-plan” step.
- Task Decomposition: Decomposition-first methods identify sub-goals before planning for them successively, while interleaved methods reveal and plan one or two sub-tasks at each state.Decomposition-first methods include HuggingGPT, Plan-and-Solve, and ProgPrompt; interleaved methods include CoT, ReAct, PAL, PoT, and Visual ChatGPT.
- Decomposition-First Methods: HuggingGPT uses an LLM controller to decompose multimodal tasks, assign dependencies, select models, and generate final responses.Its model ecosystem includes image, video, and speech capabilities from the Hugging Face Hub.
- Discussions: Decomposition-first planning preserves stronger correlations with original tasks, whereas interleaved planning dynamically adapts decomposition using environmental feedback.The former reduces task forgetting and hallucinations, while the latter improves fault tolerance.
- Discussions: Task decomposition requires extra reasoning and generation, increasing time and computational costs.When tasks produce dozens of sub-tasks, limited context length can cause the agent to forget planning trajectories; excessively long trajectories can also trigger hallucinations and goal deviation.
4 Multi-Plan Selection
Multi-plan selection addresses the risk that a single LLM-generated plan is suboptimal or infeasible by generating multiple candidates and selecting among them with search or voting strategies. It broadens solution exploration but increases computation and inherits uncertainty from LLM-based plan evaluation.
- Multi-Plan Selection: Multi-plan selection generates multiple candidate plans and chooses an optimal plan because a single LLM plan may be suboptimal or infeasible.The process consists of multi-plan generation followed by optimal plan selection.
- Multi-Plan Generation: Self-consistency samples distinct reasoning paths, while Tree-of-Thought generates plans through sample and propose strategies.Sampling can use temperature or top-k decoding.
- Plan Selection: Selection strategies include majority voting, BFS, DFS, and tree-based methods in which the LLM evaluates actions before expansion.Self-consistency uses majority vote; ToT, LLM-MCTS, and RAP use tree structures.
- Discussions: Multi-plan selection explores a broader range of solutions but increases computational demands and may be costly for online services.Large token counts or computations make resource constraints especially important.
- Discussions: LLM-based plan ranking remains under scrutiny, and stochastic selection can reduce the consistency and reliability of chosen plans.The survey identifies further validation and fine-tuning as necessary for ranking capabilities.
5 External Planner-Aided Planning
External planner-aided planning combines LLMs with symbolic or neural planners to address complex constraints, infeasible actions, and planning efficiency. The LLM commonly formalizes tasks or interprets feedback, while external planners provide domain-specific planning or formal reasoning.
- External Planner-Aided Planning: External planners address difficulties that arise when LLMs face intricate constraints, mathematical problems, or the need to generate admissible actions.The survey distinguishes symbolic planners from neural planners.
- Symbolic Planners: Symbolic planners use formal models such as PDDL to identify paths from initial states to desired goal states.LLM+P converts natural-language problems into PDDL or ASP representations before symbolic solving.
- Neural Planners: Neural planners learn planning policies from collected data through reinforcement learning or imitation learning within specific domains.Examples include DRRN, which models planning as a Markov Decision Process, and Decision Transformer, which clones human decision-making behavior.
- Neural Planners: Small neural planners can be efficient but often generalize poorly to complex, infrequent problems with scarce training data.This motivates combining LLMs with lightweight neural planners.
- Discussions: In assisted strategies, LLMs mainly formalize tasks, parse feedback, and provide reasoning, while symbolic systems contribute completeness, stability, and interpretability.LLM code-generation abilities can accelerate construction of symbolic models that traditionally require substantial human expertise.
6 Reflection and Refinement
Reflection and refinement improve LLM-agent planning by using feedback from failures or evaluations to correct errors and escape thought loops. These updates can influence later plans through textual memory, but their convergence to a specified goal is not guaranteed.
- Reflection and Refinement: Reflection and refinement improve fault tolerance and error correction when hallucinations or limited reasoning cause planning errors and thought loops.Summarizing failures helps agents correct subsequent attempts.
- Methods: Self-Refine iterates through generation, feedback, and refinement, while Reflexion uses trajectory evaluation and self-reflection after errors.CRITIC adds external validation through knowledge bases and search engines.
- Discussions: Self-reflection updates the agent through textual feedback rather than parameter modification, and these feedbacks can act as short-term or long-term memory.The feedback influences subsequent planning outputs through prompts.
- Discussions: Continual reflection lacks a guaranteed convergence proof showing that an LLM agent will ultimately reach a specified goal.
7 Memory-Augumented Planning
Memory-augmented planning enhances LLM-agent planning through retrieval-based and parameter-based memories. These approaches differ in update cost, storage, retrieval dependence, and ability to retain details.
- RAG-based Memory: RAG-based memory retrieves task-relevant past experiences from external storage during planning, using text, tables, or knowledge graphs.Memories may be indexed as vectors or retrieved using recency and relevance.
- Embodied Memory: Embodied memory fine-tunes an LLM on historical experiential samples, embedding commonsense knowledge, task priors, and successes or failures into model parameters.Parameter-efficient fine-tuning can reduce the cost of updating large language models.
- Embodied Memory: Fine-tuning-based methods improve planning on challenging and unseen tasks by training on action trajectories or plan dialogues.TDT achieves better success rates on challenging ScienceWorld tasks, while AgentTuning improves performance on unseen planning tasks.
- Discussions: RAG offers real-time, low-cost external updates but depends on retrieval accuracy, whereas fine-tuning offers larger memorization capacity with costly updates and weaker fine-grained retention.The two approaches therefore trade update efficiency against parameter-based storage capacity.
- Discussions: Memory-enhanced agents show improved growth and fault tolerance, but self-generated memory remains dependent on the LLM’s generation capability.Improving weaker LLM agents through self-generated memory remains challenging.
8 Evaluation
The survey evaluates planning agents across interactive gaming and retrieval environments, emphasizing task success and other performance measures. Reflexion improves complicated tasks despite consuming about twice as many tokens as ReAct.
- Interactive Gaming Environments: Interactive gaming environments provide real-time multimodal feedback, with Minecraft commonly evaluating agents by the quantity of tools they create.Agents gather materials to create tools for obtaining more rewards.
- Information Retrieval Environments: Retrieval environments include Wikipedia-based question answering and web browsing tasks, where task success rate is usually the evaluation metric.Examples include HotPotQA, Fever, WebShop, Mind2Web, and WebArena.
- Evaluation Results: About twice the tokens compared with ReAct, Reflexion achieves promising improvements on complicated ALFWorld and ScienceWorld tasks.The comparison is presented as evidence of error-correcting capability in LLMs.
9 Conclusions and Future Directions
The survey reports rising interest in using LLMs to improve agent planning and identifies several unresolved challenges. These include hallucinations, infeasible or inefficient plans, multimodal feedback, and coarse evaluation.
- Conclusions: The survey summarizes major planning directions, compares representative methods on four benchmarks, and reports that performance increases with expenses.It frames these findings alongside significant remaining challenges.
- Challenges: LLM hallucinations can produce irrational plans, prompt-unfaithful behavior, or actions involving items absent from the environment.Prompt engineering can alleviate these issues, but the survey describes them as fundamental shortcomings of LLMs.
- Challenges: Generated plans may violate complex or uncommon constraints because LLMs can fail to account for adequate preconditions.Connecting LLMs with symbolic planning models is identified as a promising direction.
- Challenges: Existing agents greedily generate plans without evaluating efficiency, motivating additional modules that assess plan efficiency.The proposed direction targets more efficient plans rather than only task completion.
- Challenges: Text-oriented LLMs face limitations when processing real-world feedback containing images, audio, and other modalities.Future work may integrate multimodal large models with revised planning strategies.
- Challenges: Current benchmarks often evaluate only final task completion and use simplistic, rule-based environmental feedback, limiting fine-grained assessment.The survey suggests more realistic evaluation environments designed with high-intelligence models.