Travel‑planning agents often produce itineraries that become infeasible after acceptance due to flight cancellations, hotel unavailability, or attraction closures. Three main revision strategies exist: full replanning, classical plan repair, and LLM‑based travel‑agent revision. Divergent task formulations and evaluation protocols have made direct comparison difficult.
We conduct a systematic empirical study on two TREK‑derived benchmark sets: 500 single‑disruption cases (both feasible and infeasible) and 200 feasible simultaneous compound‑disruption cases. The methods compared are LLM‑Z3 full replanning, IPyHOPPER hierarchical repair, and the iTIMO local‑revision adapter, evaluated on success rate, plan stability, and computational cost.
Results show that LLM‑Z3 with Gemini achieves the highest success on compound disruptions. IPyHOPPER nearly matches this configuration’s overall success on single disruptions while preserving substantially more of the accepted itinerary on successful repairs. Both hierarchical and local repairs require fewer edits and retain more commitments than full replanning. In terms of computation, IPyHOPPER uses no LLM inference, the evaluated LLM‑Z3 adapter performs a compact single‑call inference, and the iTIMO adapter consumes markedly more tokens.
From these findings we derive practical guidelines for balancing feasibility recovery, commitment preservation, and computational cost.
Review: In resource‑constrained travel scenarios, hierarchical repair offers the best trade‑off when preserving confirmed itinerary details is paramount; for complex compound disruptions with ample compute, LLM‑Z3 remains the most reliable option.