GRASP: Generating, Revising, and Assessing for Strategic Planning with Agentic AI

Summary

GRASP is an inference-time framework that decomposes LLM planning into three context-isolated stages: GenPlan distills task-level constraints and soft guidelines into a reusable plan blueprint; RevPlan explores 2-4 strategy-specific candidate plans in separate context windows; VerPlan independently scores candidates against constraints and rubrics and selects the best. Using RunAgent as a common executor, the authors evaluate on Natural Plan Calendar Scheduling, ZebraLogic, SciBench Math and GPQA with GPT-4o/GPT-4o-mini backbones. GRASP(GPT-4o) reaches 75.4% on calendar scheduling (vs 63.0% for a planner baseline) and 61.4% on ZebraLogic (vs 30.6%), at roughly 13.5x the planner's token cost; GRASP(GPT-4o-mini) achieves comparable gains at about 1.16x a plain GPT-4o call. In interleaved dual-task scheduling, GRASP(GPT-4o) holds 45.7% while direct planners collapse to about 30%, exceeding a direct GPT-5-mini planner by 14.5 points absolute. The authors candidly report null results on GPQA and SciBench Calc/Diff, attributing them to limits planning structure cannot overcome. Key limitations: most experiments are single-run (only GPQA and part of the ablation use three runs); the calendar margin over the strongest baseline, PlanGEN BoN, is not statistically significant (p=0.33); the strategy count is a task-dependent hyperparameter; and the smaller backbone collapses on triple tasks (41.2% to 28.1%).