PADME: Procedure Aware DynaMic Execution

Executing long-horizon procedures from natural language is challenging for LLM agents due to the lack of structure in free-form instructions like recipes or business workflows, often leading to execution drift or failure. We propose Procedure Aware DynaMic Execution (PADME), a framework that autonomously transforms procedural text into executable graphs capturing dependencies and decision logic. PADME employs a two-phase approach: a Teach phase for systematic structuring and an Execute phase for dynamic, realtime inputs and environment feedback-driven traversal. This graphbased representation provides an inductive bias that reduces error accumulation and ensures scalability. Empirically, PADME achieves state-of-the-art performance on four benchmarks, including ALF- World and ScienceWorld, demonstrating that agents equipped with graph-based procedure representations offer a powerful intermediate abstraction for robust and generalizable execution.

Authors

• Deepeka Garg J. P. Morgan AI Research
• Sihan Zeng J. P. Morgan AI Research
• Annapoorani L. Narayanan J. P. Morgan AI Research
• Sumitra Ganesh J. P. Morgan AI Research
• Leo Ardon J. P. Morgan AI Research

Keywords

• Autonomous Agents
• LLMs
• Graph Representation
• Workflow Automation