GRASP introduces strategy-aware multi-stage planning to improve LLM reliability

Researchers introduce GRASP, a strategy-aware, multi-stage planning framework designed to generate high-quality natural language executable plans for complex tasks where standard Large Language Models typically degrade in reliability. The system decouples the planning pipeline into specialized modules: GenPlan for global macro-guidelines, RevPlan for exploring localized strategies, and VerPlan for independent trajectory evaluation.

• GRASP establishes a new state-of-the-art across diverse datasets, yielding accuracy gains of ~12.4% on Natural Plan Calendar Scheduling and ~30.8% on ZebraLogic compared to direct LLM planners.
• The framework completely flattens the multi-task degradation penalty observed in standard planners, achieving an absolute accuracy gain of up to 16.7% in interleaved dual-task environments.
• By isolating context and enforcing strict macro-regularization, GRASP outperforms frontier reasoning models like GPT-5-mini by a margin of 14.5%.

This approach addresses the challenge of maintaining performance as task complexity increases, offering significant improvements over direct LLM planning methods.