LLMs for Planning Tasks: Transforming AI into Intelligent Planning Assistants

2025-04-18

Today, even with the advanced level of computational technology, optimization problems plague industries from logistics to manufacturing. While large language models (LLMs) have improved many aspects of AI, their ability to handle intricate planning challenges has remained surprisingly limited—until now. A fundamentally novel concept and framework developed by MIT researchers is changing how we approach optimization by transforming LLMs from conversational tools into powerful planning assistants, working well in varied decision-making scenarios.

Planning tasks, LLM – artistic impression. Image credit: Alius Noreika / AI

The Planning Challenge: Where Traditional LLMs Fall Short

Despite their impressive capabilities in content generation, translation, and creative tasks, large language models struggle when confronted with complex planning problems that have numerous interdependent variables. Consider a coffee company managing its supply chain: sourcing beans from multiple suppliers, operating various roasting facilities, and distributing products to different retail locations—all while trying to minimize costs amid increasing demand.

Asking ChatGPT directly to solve such a problem typically yields disappointing results. These models weren’t designed to handle the computational complexity inherent in optimization problems where billions of potential choices exist.

LLMFP: Bridging Natural Language and Optimization Algorithms

MIT’s Laboratory for Information and Decision Systems (LIDS) has developed a revolutionary approach called LLM-Based Formalized Programming (LLMFP). Instead of forcing LLMs to solve planning problems directly, the framework upgrades them into intelligent intermediaries that can:

Interpret natural language descriptions of complex problems
Identify decision variables and constraints
Translate these elements into mathematical formulations
Connect with specialized optimization solvers
Verify and refine solutions before presenting them to users

“Our research introduces a framework that essentially acts as a smart assistant for planning problems. It can figure out the best plan that meets all the needs you have, even if the rules are complicated or unusual,” explains Yilun Hao, a graduate student at MIT and lead author of the research.

How LLMFP Works: A Multi-Step Reasoning Approach

The LLMFP framework operates through a methodical process that mirrors how human experts would approach optimization challenges:

Problem Analysis and Formalization

When provided with a natural language description of a problem, the LLM analyzes the scenario to determine the essential decision variables and constraints that will shape the optimal solution. This mimics how optimization experts decompose complex problems into manageable components.

Mathematical Encoding

The model then encodes these elements into a mathematical formulation that can be processed by specialized optimization solvers—powerful algorithms designed specifically for tackling combinatorial optimization problems.

Self-Assessment and Refinement

What truly sets LLMFP apart is its self-assessment capability. After formulating a solution, the framework analyzes its own work, identifies potential errors or missing constraints, and refines the approach. This creates a feedback loop that dramatically improves accuracy.

For example, if optimizing a coffee shop’s supply chain, a human intuitively knows you can’t ship a negative quantity of beans, but an LLM might miss this implicit constraint. The self-assessment module would flag this error and prompt the model to correct it.

User-Friendly Output

Once a valid solution passes the self-assessment phase, LLMFP translates the technical solution back into natural language, making complex optimization outcomes accessible to non-technical users.

Impressive Performance Metrics

The MIT team rigorously tested their framework across nine diverse planning challenges, including warehouse robot routing optimization. The results speak volumes about LLMFP’s effectiveness:

85% average success rate across all testing scenarios
More than double the performance of baseline approaches (which achieved only 39%)
Consistent performance across different LLM architectures
No need for domain-specific examples or training data

Unlike competitive approaches, LLMFP doesn’t require extensive example libraries or domain-specific training. It can tackle novel planning problems immediately, making it exceptionally versatile.

So, how do we effectively utilize these new large language models?

This new kind of framework opens possibilities for intelligent planning assistance in numerous domains where optimization is critical:

Supply chain management and logistics optimization
Manufacturing production scheduling
Airline crew and equipment scheduling
Healthcare staff and resource allocation
Warehouse management and inventory optimization
Energy grid management and distribution

“With LLMs, we have an opportunity to create an interface that allows people to use tools from other domains to solve problems in ways they might not have been thinking about before,” notes Chuchu Fan, associate professor at MIT and senior author of the research.

One of LLMFP’s most valuable features is its adaptability. The framework can be configured to work with different optimization solvers simply by adjusting the prompts fed to the LLM. This flexibility makes it suitable for diverse problem domains and organizational needs.

The system can also adapt to user preferences. If it identifies that a particular user prefers not to modify certain variables (such as travel time or budget), it can prioritize solutions that respect those preferences—creating a more personalized planning experience.

Future Directions: Beyond Text-Based Inputs

Looking forward, the MIT research team aims to expand LLMFP’s capabilities to accept image inputs alongside natural language descriptions. This enhancement would allow the framework to tackle planning challenges that are difficult to fully articulate in text alone, such as spatial routing problems or layout optimizations.

Such multimodal capability would further bridge the gap between human problem-solving and AI-assisted planning, making optimization technology accessible to an even broader range of users and use cases.

Conclusion: More Advanced Planning Tools For Everyone

The LLMFP framework represents a significant step toward facilitating access to sophisticated optimization techniques. The natural language capabilities of LLMs to interface with powerful planning algorithms make it possible to find connections between everyday users and specialized computational tools that were previously available only to experts.

If you are interested in this topic, we suggest you check our articles:

Sources: MIT

Written by Alius Noreika