Plan-over-Graph: Towards Parallelable LLM Agent Schedule

ACL ARR 2026 March Submission772 Authors

Keywords: agent, planning, graph understanding, evaluation and metrics, inference methods, automatic evaluation

Abstract:

Large Language Models (LLMs) have demonstrated exceptional abilities in task planning, enabling LLMs to dive and conquer complex agentic tasks. However, challenges related to parallel scheduling remain underexplored.
This paper introduces a novel paradigm, _plan-over-graph_, where a real-world task is first decomposed into executable subtasks to construct an abstract task graph, then the abstract graph is leveraged to generate a plan for parallel execution that minimizes overall time cost.
We design an automated and controllable pipeline to generate synthetic graphs and propose a two-stage training scheme to enhance the planning capability of complex, scalable graphs.
Experimental results show that our _plan-over-graph_ method significantly improves planning performance on both API-based LLMs and trainable open-sourced LLMs, naturally supporting parallel execution and demonstrating global efficiency by normalizing complex tasks as graphs.
Further analysis confirms the scalability of our approach with respect to textual task descriptions and increasing graph complexity.

Paper Type: Long

Research Area: LLM agents

Research Area Keywords: agent, planning, graph understanding, evaluation and metrics, inference methods, automatic evaluation

Contribution Types: NLP engineering experiment, Publicly available software and/or pre-trained models

Languages Studied: English

Submission Number: 772