Back to AI Research

AI Research

MAPLE: Memory-Augmented Planning with Language and... | AI Research

Key Takeaways

  • MAPLE: Memory-Augmented Planning with Language and Evolution introduces a new approach for AI agents to handle real-world optimization tasks that change over...
  • Domain practitioners understand their business constraints but may lack operations-research expertise or dedicated support.
  • LLM-based optimization agents translate natural-language requirements into models or solver programs that established optimization tools can execute.
  • This progress makes optimization more accessible, but real-world operations are dynamic: changing demand, resources, and priorities require updates to data, constraints, and objectives.
  • Methods centered on isolated requests offer limited support for rapid adaptation that preserves earlier decisions and reuses useful search results.
Paper AbstractExpand

Domain practitioners understand their business constraints but may lack operations-research expertise or dedicated support. LLM-based optimization agents translate natural-language requirements into models or solver programs that established optimization tools can execute. This progress makes optimization more accessible, but real-world operations are dynamic: changing demand, resources, and priorities require updates to data, constraints, and objectives. Methods centered on isolated requests offer limited support for rapid adaptation that preserves earlier decisions and reuses useful search results. We introduce MAPLE (Memory-Augmented Planning with Language and Evolution), an agent for maintaining optimization problems through successive natural-language requests. MAPLE combines language-based problem construction with mathematical programming and evolutionary search. It retains the optimization program, accepted plans, earlier updates, and candidate solutions for subsequent requests. We introduce NLDO, a benchmark of 15 trajectories and 180 updates spanning selection, scheduling, rostering, routing, and cloud-resource placement. In the main evaluation, MAPLE completes all trajectories and achieves online scalar quality of 0.951 and a Pareto hypervolume ratio of 0.875. Controlled comparisons further show that maintaining executable state improves update validity and can preserve useful search information across substantial revisions.

MAPLE: Memory-Augmented Planning with Language and Evolution introduces a new approach for AI agents to handle real-world optimization tasks that change over time. While existing LLM-based tools are good at creating initial plans, they often struggle when business needs evolve—such as when new orders arrive, resources become unavailable, or priorities shift. MAPLE addresses this by maintaining a persistent "executable state" that allows the agent to remember past decisions, reuse previous search results, and adapt to new instructions without starting from scratch.

How MAPLE Works

MAPLE functions as an intelligent workbench that keeps track of the entire optimization process. It uses three core components to manage updates:

  • Typed Search-Space Scaffolding (TSS): This provides a structured framework that binds the AI's decisions to specific, reliable mathematical solvers and evolutionary search tools. By defining how to represent and evaluate problems, it ensures the agent stays within valid operational boundaries.

  • Live Problem Decomposition (LPD): When a user provides a new request, this component identifies exactly which parts of the existing model, data, or constraints need to change, leaving the rest of the system intact.

  • Live State Memory (LSM): This acts as a library for the agent, storing accepted plans and previous search candidates. If a user asks to keep a specific vehicle assignment from a previous plan, the agent can retrieve that information and enforce it during the next round of optimization. The same ai evaluation question is explored in DRACO, which adds a research perspective.

Intelligent Adaptation

A key feature of MAPLE is its ability to decide how to perform a new search. After a change is requested, a "restart selector" analyzes the update to determine if it should perform a "Warm" start—reusing previous solutions to save time and maintain consistency—or a "Full" start, which clears the memory to begin fresh. This decision is based on a verification process that checks if the new constraints make old solutions obsolete or if they can still provide a useful starting point for the new plan. The same ai evaluation question is explored in JarvisGUI, which adds a research perspective.

Performance and Results

The authors tested MAPLE using a new benchmark called NLDO, which consists of 15 complex planning trajectories and 180 natural-language updates covering tasks like scheduling, routing, and cloud-resource placement. In these evaluations, MAPLE successfully completed every trajectory. It achieved an online scalar quality score of 0.951 and a Pareto hypervolume ratio of 0.875, significantly outperforming other tested methods. The results demonstrate that maintaining an executable state is critical for ensuring that an agent remains reliable and consistent as it adapts to ongoing, real-world operational changes. The same ai evaluation question is explored in COBRA-Skills, which adds a research perspective. as detailed in the full paper on Arxiv

More from Franklin

COBRA-SkillsresearchWhat Should an Agent Forget? Separating...researchDRACOresearchJarvisGUIresearchWhen Agents DisagreeresearchEvoSCMresearch

Comments (0)

No comments yet

Be the first to share your thoughts!