GRASP: Generating, Revising, and Assessing for Strategic Planning with Agentic AI
GRASP splits complex planning into separate generate, revise, and verify stages to keep LLM plans from collapsing under harder workloads.
The paper says the framework uses GenPlan for macro-guidelines, RevPlan for isolated local strategy exploration, and VerPlan for independent multi-criteria evaluation. In tests, it reports gains over direct LLM planners on calendar scheduling, ZebraLogic, and SciBench Math. The authors also say GRASP removes the usual multi-task degradation penalty and beats GPT-5-mini by 14.5% under their setup. Accepted at REALM at EMNLP 2026. ArXiv · AI/CL/LG's note
The paper says the framework uses GenPlan for macro-guidelines, RevPlan for isolated local strategy exploration, and VerPlan for independent multi-criteria evaluation. In tests, it reports gains over direct LLM planners on calendar scheduling, ZebraLogic, and SciBench Math. The authors also say GRASP removes the usual multi-task degradation penalty and beats GPT-5-mini by 14.5% under their setup. Accepted at REALM at EMNLP 2026. ArXiv · AI/CL/LG's note
score 5