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

Problem

The paper addresses the challenge of generating high-quality natural language executable plans for complex tasks, a gap in the current literature on planning with AI. The work is presented as a preprint and has not undergone peer review.

Method

The authors propose GRASP, a strategy-aware, multi-stage planning framework composed of three main components:

• GenPlan: This component pre-compiles global macro-guidelines to inform the planning process.
• RevPlan: It explores alternative localized strategies within isolated context windows, allowing for flexibility in planning.
• VerPlan: This component independently evaluates trajectories using a multi-criteria discriminator, ensuring that the generated plans meet various criteria for effectiveness.

Results

The GRASP framework demonstrates notable improvements over direct LLM planners in several benchmarks:

• Natural Plan Calendar Scheduling: Achieved a 12.4% improvement compared to direct LLM planners.
• ZebraLogic: Showed a 30.8% improvement over direct LLM planners.
• SciBench Math: The available text does not report quantitative results.
• Multi-task scaling: The framework flattens the multi-task degradation penalty, indicating enhanced performance across multiple tasks.
• Interleaved dual-task environments: Achieved a 16.7% absolute accuracy gain compared to direct LLM planners.
• Frontier reasoning models (GPT-5-mini): GRASP achieved a 14.5% improvement over this baseline.

Limitations

The authors do not report any limitations in their work. However, the absence of quantitative results for the SciBench Math benchmark may indicate areas for further exploration or validation.

Why it matters

The implications of this work are significant for downstream applications in AI planning and decision-making. By improving the quality and effectiveness of natural language executable plans, GRASP could enhance the capabilities of AI systems in complex task execution, potentially leading to more efficient and reliable AI-driven solutions in various domains.