often degrade sharply as task complexity increases. GRASP addresses this by decoupling the planning pipeline into three specialized modules: GenPlan pre-compiles global macro-guidelines, RevPlan explores alternative localized strategies in isolated context windows, and VerPlan evaluates trajectories with a multi-criteria discriminator. This structure prevents cross-contamination and error cascades.
Evaluated across diverse datasets, GRASP achieved substantial accuracy gains over direct LLM planners: about 12.4% on Natural Plan Calendar Scheduling, 30.8% on ZebraLogic, and improvements on SciBench Math. In interleaved dual-task environments, it completely flattened the multi-task degradation penalty, achieving an absolute accuracy gain of up to 16.7% over direct planners and outperforming frontier reasoning models like GPT-5-mini by 14.5%.
The key interest is that GRASP demonstrates that strict contextual isolation and macro-regularization can systematically overcome the “Curse of Instructions” without requiring more powerful models. This suggests a viable path for complex, long-horizon planning tasks in enterprise and agentic AI systems.