Scope: Hierarchical planner beats LLMs, 55x faster, 1/160k size (skyfall.ai)

SCOPE (Subgoal-COnditioned Pretraining for Efficient planning) has emerged as a groundbreaking alternative to large language models (LLMs) for task planning, achieving a 56% success rate, which surpasses the baseline ADaPT's 52%. What sets SCOPE apart is its efficiency; it operates 55 times faster than ADaPT, completing tasks in just three seconds on a single NVIDIA A10 GPU, and boasts a significantly smaller model size with only 11 million parameters—160,000 times smaller than larger counterparts like GPT-4o. More importantly, SCOPE incurs zero ongoing API costs after an initial one-time setup, making it an attractive, practical option for real-world applications. The innovative design of SCOPE involves a hierarchical structure where a manager agent proposes subgoals while an employee agent executes them. This system relies on a one-time knowledge extraction from an LLM to generate decomposition functions that can handle extensive datasets without additional reliance on LLM queries. Through a combination of imitation learning and reinforcement learning, SCOPE teaches the manager to adapt to the employee’s strengths and weaknesses, enhancing its overall task completion success. This research not only demonstrates the potential of SCOPE to outperform traditional LLM-based planning but also reveals that even suboptimal LLM-generated subgoals can significantly elevate performance, underscoring the efficiency of hierarchical planning systems.