Language Models as One-Time Teacher for Hierarchical Planning in Text Environs (arxiv.org)

Researchers have introduced SCOPE (Subgoal-COnditioned Pretraining for Efficient planning), a novel one-shot hierarchical planning framework designed to enhance efficiency in long-term planning within complex, text-based environments. Traditional methods often require extensive querying of large language models (LLMs) during both training and inference, leading to high computational costs and inefficiencies. SCOPE addresses these challenges by generating subgoals from example trajectories only at initialization, allowing a lightweight student model to be pretrained without needing repetitive LLM calls. This approach not only improves computational efficiency but also enables faster inference times, reducing the duration from 164.4 seconds to just 3.0 seconds, while achieving a success rate of 0.56 in planning tasks. The significance of SCOPE lies in its ability to streamline hierarchical planning by leveraging LLM-generated subgoals more effectively, even if these subgoals are not always optimal. This advancement represents a shift in how LLM-driven systems can be deployed in practical scenarios, emphasizing the potential for lightweight models that still harness the semantic knowledge of LLMs. By providing a strong foundation for hierarchical goal decomposition in text-based tasks, SCOPE could pave the way for future research and applications that require rapid, efficient decision-making in complex environments.