When your flight gets canceled, should an AI travel assistant start from scratch or just patch the plan?
A new study compares three approaches to fixing travel itineraries wrecked by canceled flights, sold-out hotels, or shuttered attractions: full replanning with an LLM-Z3 system, hierarchical plan repair via IPyHOPPER, and a local-revision LLM adapter called iTIMO. Researchers tested them on 500 single-disruption cases and 200 cases with multiple simultaneous disruptions, all derived from the TREK travel-planning benchmark. LLM-Z3 paired with Gemini won on the hardest compound-disruption scenarios. But IPyHOPPER came close on single disruptions and kept far more of the original itinerary intact, all without calling an LLM at all.
That gap matters for anyone building a real travel agent, not just a research demo. Full replanning can quietly cancel bookings and preferences a traveler already locked in, and it costs real inference money every time a flight gets delayed. A repair method that fixes what broke and leaves everything else alone is cheaper to run and less likely to annoy a customer who just wanted a new hotel, not a new vacation.
Full AI replanning still wins when several things go wrong at once, but for the everyday case of one canceled flight, the fanciest model in the room is not necessarily the one that should touch your itinerary.