Research map / Multi-agent systems and games / LLM agents
ATLAS: Constraints-Aware Multi-Agent Collaboration for Real-World Travel Planning
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes ATLAS, a multi-agent framework (Search, Constraint Manager, Planner, Checker, Search Advisor) that discovers explicit and implicit constraints and adaptively searches to fix information gaps for travel planning, evaluated on TravelPlanner, Flex-TravelPlanner, and a live-search multi-turn setting.
Problem
LLM agents struggle to produce valid travel plans when they must both search for information and discover constraints (explicit and implicit) without prior knowledge, especially under multi-turn evolving user constraints.
Approach
Formalizes travel planning as a dynamic CSP whose domains are built from live tool calls; a Constraint Manager codifies constraints, a Planner-Checker loop iteratively revises plans, and a Search Advisor diagnoses unsatisfiable cases to direct new searches, with cached domains reused across turns.
Key results
- On TravelPlanner test set with Gemini-2.5-Pro, ATLAS reaches 35.00% final pass vs 22.70% for ReAct+Reflexion and 19.50% for ReAct.
- In live multi-turn search after three turns, ATLAS final pass rises from 26.1% to 83.9%, versus 58.9% for ReAct and 27.2% for the monolithic agent.
- On Flex-TravelPlanner 3-turn settings, ATLAS roughly doubles ReAct's final pass rate (e.g., 33.60% vs 15.34%).
Surprising findings
- Adding search to advanced planners (Reflexion, EvoAgent) yields no benefit, and ReAct+EvoAgent underperforms plain ReAct on the test set.
- On live search, the gap between ATLAS and ReAct is modest in the first turn, and gains appear mainly through multi-turn feedback.
Limitations
- Relies on benchmark-specific unconventional hints for the largest gains; without them final pass is lower (44% vs 60% on validation).
- Live search evaluation uses an LLM judge with the same base model for hallucination checking.
- Compute and cost trade-offs of many planner-checker and search iterations may be high.
Experimental setup
Datasets: TravelPlanner Flex-TravelPlanner
Benchmarks: TravelPlanner Flex-TravelPlanner TravelPlanner with live Google Search
Metrics: delivery rate micro pass rate macro pass rate final pass rate no hallucination rate
Baselines: ReAct ReAct+Reflexion ReAct+EvoAgent PMC Monolithic agent
Builds on: ReAct Reflexion EvoAgent PMC TravelPlanner Flex-TravelPlanner
Closest papers by meaning
- WorldTravel: A Realistic Multimodal Travel-Planning Benchmark with Tightly Coupled Constraints (ICML 2026 · similarity 0.91)
- Wide-Horizon Thinking and Simulation-Based Evaluation for Real-World LLM Planning with Multifaceted Constraints (NeurIPS 2025 · similarity 0.90)
- HiMAP-Travel: Hierarchical Multi-Agent Planning for Long-Horizon Constrained Travel (ICML 2026 · similarity 0.89)
- TravelPlanner: A Benchmark for Real-World Planning with Language Agents (ICML 2024 · similarity 0.89)
- TRIP-Bench: A Benchmark for Long-Horizon Interactive Agents in Real-World Scenarios (ICML 2026 · similarity 0.87)
- SPIRAL: Symbolic LLM Planning via Grounded and Reflective Search (AAAI 2026 · similarity 0.85)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.