Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Beyond Entangled Planning: Task-Decoupled Planning for Long-Horizon Agents

Yunfan Li, Bingbing Xu, Xueyun Tian, Xiucheng Xu, Huawei Shen

TL;DR

This work identifies entangled planning as a key bottleneck for long-horizon, LLM-driven agents and introduces Task-Decoupled Planning (TDP), a modular framework that decouples global task structure from local decision making. A global Supervisor builds a DAG of sub-tasks and schedules nodes, while a Planner/Executor pair operates with node-scoped contexts to limit reasoning and replanning to the active sub-task, augmented by Self-Revision to maintain a feasible, up-to-date graph. Empirical results on TravelPlanner, HotpotQA, and ScienceWorld show that TDP achieves strong performance across diverse task types while drastically reducing token usage compared to baselines, demonstrating improved robustness and efficiency. The paper argues that explicit task decoupling can balance global coherence with local adaptability, offering a practical design principle for reliable long-horizon autonomous agents.

Abstract

Recent advances in large language models (LLMs) have enabled agents to autonomously execute complex, long-horizon tasks, yet planning remains a primary bottleneck for reliable task execution. Existing methods typically fall into two paradigms: step-wise planning, which is reactive but often short-sighted; and one-shot planning, which generates a complete plan upfront yet is brittle to execution errors. Crucially, both paradigms suffer from entangled contexts, where the agent must reason over a monolithic history spanning multiple sub-tasks. This entanglement increases cognitive load and lets local errors propagate across otherwise independent decisions, making recovery computationally expensive. To address this, we propose Task-Decoupled Planning (TDP), a training-free framework that replaces entangled reasoning with task decoupling. TDP decomposes tasks into a directed acyclic graph (DAG) of sub-goals via a Supervisor. Using a Planner and Executor with scoped contexts, TDP confines reasoning and replanning to the active sub-task. This isolation prevents error propagation and corrects deviations locally without disrupting the workflow. Results on TravelPlanner, ScienceWorld, and HotpotQA show that TDP outperforms strong baselines while reducing token consumption by up to 82%, demonstrating that sub-task decoupling improves both robustness and efficiency for long-horizon agents.

Beyond Entangled Planning: Task-Decoupled Planning for Long-Horizon Agents

TL;DR

This work identifies entangled planning as a key bottleneck for long-horizon, LLM-driven agents and introduces Task-Decoupled Planning (TDP), a modular framework that decouples global task structure from local decision making. A global Supervisor builds a DAG of sub-tasks and schedules nodes, while a Planner/Executor pair operates with node-scoped contexts to limit reasoning and replanning to the active sub-task, augmented by Self-Revision to maintain a feasible, up-to-date graph. Empirical results on TravelPlanner, HotpotQA, and ScienceWorld show that TDP achieves strong performance across diverse task types while drastically reducing token usage compared to baselines, demonstrating improved robustness and efficiency. The paper argues that explicit task decoupling can balance global coherence with local adaptability, offering a practical design principle for reliable long-horizon autonomous agents.

Abstract

Recent advances in large language models (LLMs) have enabled agents to autonomously execute complex, long-horizon tasks, yet planning remains a primary bottleneck for reliable task execution. Existing methods typically fall into two paradigms: step-wise planning, which is reactive but often short-sighted; and one-shot planning, which generates a complete plan upfront yet is brittle to execution errors. Crucially, both paradigms suffer from entangled contexts, where the agent must reason over a monolithic history spanning multiple sub-tasks. This entanglement increases cognitive load and lets local errors propagate across otherwise independent decisions, making recovery computationally expensive. To address this, we propose Task-Decoupled Planning (TDP), a training-free framework that replaces entangled reasoning with task decoupling. TDP decomposes tasks into a directed acyclic graph (DAG) of sub-goals via a Supervisor. Using a Planner and Executor with scoped contexts, TDP confines reasoning and replanning to the active sub-task. This isolation prevents error propagation and corrects deviations locally without disrupting the workflow. Results on TravelPlanner, ScienceWorld, and HotpotQA show that TDP outperforms strong baselines while reducing token consumption by up to 82%, demonstrating that sub-task decoupling improves both robustness and efficiency for long-horizon agents.
Paper Structure (28 sections, 4 figures, 1 table, 1 algorithm)

This paper contains 28 sections, 4 figures, 1 table, 1 algorithm.

Table of Contents

  1. Introduction
  2. Related Work
  3. LLM-based Agent Planning
  4. Task Decomposition
  5. Task Decomposition.
  6. Method
  7. Motivation
  8. Supervisor: Global Decomposition and Dependency-Aware Scheduling
  9. Planner and Executor: Local Planning, Acting, and Local Repair
  10. Self-Revision: Graph Maintenance and Specification Refinement
  11. Experiment
  12. Experimental Setup
  13. Benchmarks
  14. Evaluation Metrics
  15. Baselines
  16. ...and 13 more sections

Figures (4)

  • Figure 1: Comparison between step-wise planning (a), on-shot planning (b) and TDP (c) on a TravelPlanner task.
  • Figure 2: Overview of TDP. The Supervisor decomposes the task into a dependency graph; the Planner & Executor solve each decoupled sub-task node; Self-Revision updates the graph after execution.
  • Figure 3: Cost comparison on HotpotQA and ScienceWorld: average output tokens (left axis) and performance (right axis; delivery accuracy / average reward) for Plan-and-Act vs TDP under DeepSeek-V3.2 and GPT-4o.
  • Figure 4: Illustration of TDP task decomposition, node-local planning and replanning and isolated context on a TravelPlanner case.