Inroads to personalized AI trip planning

2025-06-10 · Source: MIT News - Natural language processing · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Robotics & Autonomous Systems · Depth: Advanced, medium

Summary

A new framework from the MIT-IBM Watson AI Lab enhances large language models (LLMs) to effectively plan complex travel itineraries. Traditional LLMs struggle with multi-constraint logistical problems, achieving viable solutions less than 4 percent of the time, even with tools. This new method combines LLMs (tested with GPT-4, Claude-3, Mistral-Large) with a satisfiability modulo theories (SMT) solver and algorithms. The LLM acts as a translator, converting natural language requests into a problem format the solver can process. The solver rigorously checks constraints and identifies conflicts, allowing the system to propose alternatives to the user. The framework achieved over a 90 percent pass rate on the TravelPlanner dataset, significantly outperforming baselines, and demonstrated robustness with unseen constraints and paraphrased queries, achieving 78.6-91.7 percent success on new datasets.

Key takeaway

For AI Scientists developing planning systems, this research demonstrates that integrating LLMs with formal verification tools like SMT solvers dramatically improves success rates for complex, multi-constraint problems. You should consider this hybrid architecture to overcome LLM limitations in rigorous logistical and mathematical reasoning, ensuring verifiable and robust solutions in domains like travel planning, robotics, or task allocation.

Key insights

Combining LLMs with formal solvers enables robust, verifiable solutions for complex, multi-constraint planning problems.

Principles

• LLMs excel at translation, not rigorous optimization.
• Solvers provide certifiable constraint satisfaction.
• Hybrid systems overcome individual LLM/solver limitations.

Method

The method involves an LLM parsing user prompts into planning steps, converting these into executable Python code that calls APIs and an SMT solver, and then iteratively refining plans based on solver feedback.

In practice

• Integrate LLMs with SMT solvers for complex planning.
• Use LLMs for natural language translation to formal systems.
• Apply this hybrid approach to logistics, robotics, and task allocation.

Topics

• Personalized AI Trip Planning
• Large Language Models
• Satisfiability Modulo Theories
• Combinatorial Optimization
• Formal Verification

Best for: AI Scientist, Research Scientist, AI Researcher, AI Engineer, AI Data Scientist

Related on AIssential

• Generative Floor Plan Design with LLMs via Reinforcement Learning with Verifiable Rewards — LLMs can generate highly constrained, structured floor plans by combining supervised fine-tuning with reinforcement learning and verifiable rewards.
• Formalizing Mathematics at Scale — AutoformBot's multi-agent system makes large-scale autoformalization of graduate mathematics economically and technically feasible.
• Interpretable and Verifiable Hardware Generation with LLM-Driven Stepwise Refinement — LLMs can generate verifiable hardware designs by integrating formal methods and stepwise refinement.
• Reliable Reasoning with Large Language Models via Preference-Based Maximum Satisfiability — LLMs can achieve reliable, verifiable optimization by externalizing reasoning into MaxSAT code generation and solver-based verification.
• Agentic AI for Trip Planning Optimization Application — Orchestrated agentic AI significantly improves trip planning optimization and robustness over single-agent or uncoordinated multi-agent systems.
• Can AI Reason Like an Urban Planner? Benchmarking Large Language Models Against Professional Judgment — The UPBench framework reveals LLMs struggle with context-sensitive, lower-order urban planning tasks despite performing well on higher-order analysis.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by MIT News - Natural language processing.