acceptodds
Under review as a conference paper at ICLR 2027

Coordinating Domain-Specialized LLM Agents for Multi-City Travel Planning

Abstract

Multi-city travel planning is a tightly coupled, multi-scale decision problem requiring mobility, accommodation, and activity choices to remain jointly feasible under temporal, spatial, budget, and user-specific constraints. We propose a structured agentic framework that explicitly coordinates these interdependent decisions while separating semantic reasoning from deterministic execution. Domain-specialized LLM agents select from grounded candidates and encode their choices, alternatives, evidence, and unresolved dependencies as decision contracts. A Cross-Domain Coordinator first provides global planning guidance and then reconciles these contracts into a joint contract under explicit authority constraints. The joint contract bounds downstream city-day composition and solver-assisted temporal-spatial scheduling, while deterministic controls validate decisions before execution and verify the materialized itinerary afterward. Detected violations are routed to the responsible decision scope for bounded revision and re-execution. The framework achieves Final Pass Rates of 74.40% on TravelPlanner and 59.44% on ChinaTravel. Ablations show that cross-domain coordination is critical for coupled multi-city decisions, whereas specialist LLM reasoning is most beneficial for semantically rich city-day planning.

Primary area:
applications to robotics, autonomy, planning
Keywords:
Multi-city travel planning, LLM agents

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_rHeBOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The authors propose an agentic framework that keeps grounded travel decisions verifiable through selection, construction, andrevision. The framework achieves a 74.40% Final Pass Rate on the official TravelPlanner Test1000 and 59.44% on a study-constructed ChinaTravel Test180 subset.

Critical strengths
  1. The research scenario is meaningful and interesting, using multi-agent collaboration for travel planning.
  2. The design of the Cross-Domain Coordinator is effective and achieves good experimental results.
Critical weaknesses
  1. The core method has limited novelty. Although it focuses on a specific application scenario, it does not involve any LLM training and its innovation is limited to the level of prompt engineering.
  2. In the ablation study, removing the Specialist actually improves performance, raising questions about its necessity.
  3. Some of the paper's experimental results show only marginal improvements over the state of the art.
Questions for authors

The same as weaknesses.

Reviewer_s4wAOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes a multi-agent framework that decouples high-level LLM reasoning from deterministic scheduling to satisfy complex temporal, spatial, and budget constraints. Its primary contributions are a cross-domain coordination mechanism -- where specialized agents reconcile choices through structured decision contracts -- and a two-stage construction and verification pipeline that pairs an OR-Tools solver with automated monitors for bounded plan repair. Experiments on the TravelPlanner and ChinaTravel benchmarks validate the framework, achieving state-of-the-art Final Pass Rates of 74.40% and 59.44%, respectively, while ablation studies confirm the necessity of both coordination and specialist reasoning.

Critical strengths

The paper's primary strength is its hybrid architecture, which uses structured decision contracts to coordinate domain-specialized LLM agents before delegating fine-grained scheduling and constraint verification to deterministic solvers and monitors. This design is consequential to the overall assessment because it directly resolves a core failure mode of multi-agent LLM planners -- hallucinated feasibility and cascading cross-domain conflicts -- driving a state-of-the-art improvement of over 21 percentage points on TravelPlanner and demonstrating the practical value of structured symbolic coordination.

Critical weaknesses
  • The paper’s own ablations undermine the necessity of its full multi-agent architecture. On TravelPlanner, replacing the domain-specialized LLM agents with simple deterministic rules actually improves the Final Pass Rate while avoiding major token and latency overhead. Conversely, on ChinaTravel, where specialist LLM reasoning helps, the paper's other core mechanisms -- cross-domain coordination, evidence sharing, and bounded repair -- provide no statistically significant gains, and the full system fails to beat the existing NeSy Planning baseline on the full dataset. This is consequential to the overall assessment because the proposed components are never jointly justified: they are either outperformed by simple heuristics or ineffective where heuristics fall short.

  • Rather than offering a generalizable planning paradigm, much of the framework’s constraint satisfaction relies on manual, benchmark-specific engineering. Both the deterministic monitors/solvers and the agent prompts explicitly hard-code evaluation protocols and custom solver penalties tailored to TravelPlanner and ChinaTravel. This is consequential to the overall assessment because it blurs the line between genuine multi-agent reasoning and a rule-based pipeline, limiting the method's scientific insight and generalizability to new domains.

  • The baseline comparisons (e.g., against ATLAS) do not control for the underlying LLM backbone, leaving it unclear whether the performance gains stem from the proposed framework or from using a stronger base model (such as DeepSeek V4 Flash). This is consequential to the overall assessment because, without evaluating across multiple backbones or matching baseline model settings, the true empirical contribution of the proposed coordination architecture cannot be reliably isolated.

Questions for authors
  • Since deterministic rules outperform the LLM domain specialists on TravelPlanner while coordination, evidence sharing, and bounded repair yield no significant gains on ChinaTravel, under what conditions are both components simultaneously necessary? Can you show a setting where the full architecture jointly outperforms simple heuristics and baselines like NeSy Planning?

  • How does performance change if the benchmark-specific prompt constraints, evaluation handlers, and custom solver penalties are replaced with a generic constraint specification, and what manual effort is required to adapt the pipeline to a new domain?

  • To isolate the framework's contribution from base model capability, how does your method perform when using the exact same LLM backbones as prior baselines (e.g., ATLAS), or when those baselines are evaluated with DeepSeek V4 Flash?

Reviewer_hFroOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper addresses multi-city itinerary planning under multiple constraints. The challenge is that transportation, accommodation, and activity choices interact, and conflicts revealed during daily scheduling may require revisiting earlier decisions. The authors use four domain specialists and a coordinator to reconcile choices through decision contracts, followed by scheduling, verification, and revision. The framework achieves 74.40% FPR on the official TravelPlanner Test1000 and 59.44% on a ChinaTravel Test180 subset constructed for this study.

Critical strengths
  1. **The framework has a sensible overall organization:**Different specialists handle city, transportation, accommodation, and activity choices, while coordination, scheduling, and verification connect their decisions. Responsibilities are clearly laid out, and the feedback process provides a practical way to revise problematic choices.

  2. **The ablations offer useful evidence for cross-domain coordination: **removing it reduces TravelPlanner validation FPR from 72.78% to 47.78%. The different specialist-selection results across benchmarks also help illustrate where LLM reasoning may be beneficial.

Critical weaknesses
  1. A comparison with a planning agent that dynamically delegates to subagents would help clarify the value of fixed specialists and decision contracts. The stronger deterministic-selection result on TravelPlanner validation (86.67% vs. 72.78%) also deserves discussion.

  2. Evaluation primarily uses one DeepSeek backbone. Testing other models, such as GPT or Claude, would strengthen evidence for generality.

  3. Given median workloads of 18–21 calls and roughly 95,000–108,000 tokens per request, cost-matched comparisons with CoT and subagent baselines would help assess the additional computation and suitability for interactive use.

Questions for authors
  1. Could the authors clarify the advantage of the fixed four-specialist architecture over dynamic task decomposition, and identify which domain-selection steps benefit most from LLM reasoning?

  2. How does the framework compare with alternative backbones and simpler workflows under comparable information access and inference budgets, considering both FPR and end-to-end cost?

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.