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

Problem

The paper addresses a significant gap in the capability of planner-executor systems, specifically the distinction between plan selection and execution failures. It highlights the need for a better understanding of how large language models (LLMs) can effectively declare and execute plans, which is crucial for improving the reliability of AI agents in complex tasks. The work is presented as a preprint, indicating it has not yet undergone peer review.

Method

The authors propose a novel framework termed Planning-as-Routing, where the LLM declares one of four planning modes: Predefined, Sequential, Hierarchical, or Search. A deterministic router is employed to dispatch tasks to corresponding pattern-specific executors based on the declared mode. This architecture aims to enhance the execution fidelity of the plans by ensuring that the appropriate executor is utilized for each task type, thereby addressing the identified gap in execution reliability.

Results

The paper reports several key performance metrics:

• Plan Preservation Rate: Ranges from 22% to 45% when compared to a generic Plan+ReAct approach across three benchmarks, indicating varying degrees of effectiveness in maintaining the integrity of the declared plans.
• Task Success Improvement on ALFWorld: The success rate improves from 0.48 to 0.92 when using pattern-specific executors, demonstrating a significant enhancement in task execution.
• Task Success Improvement on SWE-bench: The success rate increases from 0.36 to 0.44, again showcasing the benefits of employing pattern-specific executors over generic methods.

These results suggest that the proposed routing mechanism can lead to substantial improvements in task execution success rates.

Limitations

The authors acknowledge several limitations in their work:

• Current LLMs do not consistently select the most effective planning mode for each task, which can lead to suboptimal execution outcomes.
• While few-shot examples can enhance mode selection in certain benchmark-model combinations, this improvement is not universally applicable across all scenarios, indicating a need for further refinement in the selection process.

Why it matters

This research has significant implications for the development of more reliable AI agents capable of executing complex plans. By improving the fidelity of plan execution through a structured routing mechanism, the findings could inform future work on enhancing LLM capabilities in real-world applications, particularly in domains requiring high levels of task specificity and execution accuracy. The insights gained from this study may also contribute to the broader understanding of how to effectively integrate planning and execution in AI systems.