Do LLM Agents Execute the Plans They Declare? From Planning-Mode Declaration to Pattern-Specific Execution

Imagine you're building an AI that needs to plan and execute tasks over a long period, like managing a project or navigating a complex environment. The challenge lies in ensuring that the AI not only chooses the right plan but also follows through effectively. Current systems often struggle because they can't clearly separate the planning phase from execution, leading to failures that are hard to diagnose — this is what's called the Plan Declaration–Execution Gap. When an AI generates a plan, it might not stick to it during execution, especially if the plan is lengthy or complex. This can result in a lack of coherence in the actions taken, which is frustrating when you're relying on the AI to get things done reliably.

To address this, the authors propose a method called Planning-as-Routing, where the AI first declares a specific planning mode — like Predefined, Sequential, Hierarchical, or Search — and then a router directs the task to the appropriate executor designed for that mode. Their findings show that using these pattern-specific executors significantly improves task success rates, especially in environments like ALFWorld and SWE-bench. While current LLMs still struggle with selecting the best planning mode for each task, the introduction of few-shot examples helps in some cases. This means that for anyone building AI systems, understanding how to effectively route tasks based on planning modes can lead to more reliable and successful outcomes.

Introduces a new framework for planning and execution in LLMs.

Evaluated across multiple benchmarks with consistent findings.