TLDR: PADME (Procedure Aware DynaMic Execution) is an AI agent framework developed by J.P. Morgan AI Research that enables large language models (LLMs) to autonomously execute long and complex procedures from natural language instructions. It works in two phases: a Teach phase, which automatically converts free-form text into an executable, graph-based representation called a ‘decision graph,’ and an Execute phase, which dynamically follows this graph, adapting to real-time inputs and environment feedback. This structured approach significantly reduces error accumulation and improves reliability and generalizability across diverse tasks like recipes, scientific experiments, and business workflows, achieving state-of-the-art performance on various benchmarks.
Intelligent agents, especially those powered by large language models (LLMs), face a significant hurdle when it comes to understanding and executing complex, multi-step procedures described in everyday language. Think of following a detailed recipe, conducting a scientific experiment, or navigating a business workflow – these tasks often involve many steps, conditional decisions, and a wide range of actions. Current LLM agents tend to struggle with these ‘long-horizon’ tasks, often losing track, accumulating errors, or failing to maintain a coherent plan.
To address this challenge, researchers at J.P. Morgan AI Research have introduced a novel agent framework called Procedure Aware DynaMic Execution, or PADME. This innovative system is designed to help AI agents autonomously execute lengthy and intricate procedures by transforming free-form instructional text into a structured, graph-based representation.
How PADME Works: A Two-Phase Approach
PADME operates through a clever two-phase methodology: the Teach phase and the Execute phase.
The Teach phase is where the magic of structuring happens. In this stage, PADME takes unstructured procedural text – like a standard operating procedure (SOP) or a manual – and converts it into an ‘executable decision graph.’ This graph is essentially a visual blueprint of the procedure, where each step is a ‘node’ and the connections between them (‘edges’) represent logical or temporal dependencies. Crucially, it explicitly captures decision points, allowing the agent to make choices based on conditions. This automated graph construction is a key differentiator, as many previous approaches required manual creation of such structures.
Once the decision graph is built and validated (potentially by human experts for critical domains), each node is enriched with executable logic, such as code snippets or API calls. This means the procedure is now ‘AI-ready,’ transformed from human-readable text into a machine-executable plan.
The Execute phase then takes this AI-ready decision graph and uses it to perform tasks dynamically. When a new task is given, the execution agent traverses the graph, following the prescribed actions. When it encounters a ‘decision node,’ it pauses and uses real-time context – like previous outputs, environmental conditions, or user input – to choose the appropriate path forward. This adaptive execution ensures consistency while allowing for flexibility in real-world scenarios.
The Benefits of Structured Reasoning
The graph-based representation offers several significant advantages. By explicitly mapping out dependencies and decision points, PADME provides an ‘inductive bias’ that dramatically reduces the chance of errors accumulating over long sequences of actions. Instead of an LLM having to search through a vast, unstructured space of possibilities, its execution is constrained to valid paths defined by the graph. This makes the execution process more reliable and predictable.
Furthermore, the separation of the Teach and Execute phases ensures quality and scalability. Expert knowledge can be encoded once into a stable logical representation, which can then be reliably reused across various contexts and even adapted to new environments or tool sets without needing to restructure the entire procedure.
Also Read:
- COMPASS: A New Framework for Enhanced AI Agent Reasoning in Complex Tasks
- Unpacking AI Agent Performance: A New Evaluation Framework
Real-World Performance
PADME has been rigorously tested across four diverse benchmarks, including the interactive environments ALFWorld and ScienceWorld, a business process corpus, and a consolidated recipe book. The results show that PADME achieves state-of-the-art performance, consistently outperforming other leading methods like Act-Only, ReAct, Chain-of-Thought, and SPRING.
Notably, PADME demonstrated strong robustness when faced with ambiguous tasks, such as inferring a dish from a list of ingredients in the ‘hard’ category of the Recipe dataset. This highlights its ability to handle real-world uncertainties more effectively than baselines.
This research marks a significant step towards creating more robust and generalizable autonomous agents capable of reliably executing complex, long-horizon procedures from natural language instructions. You can read the full paper here.
