Research map / Large language models / Agents, code and math

Wide-Horizon Thinking and Simulation-Based Evaluation for Real-World LLM Planning with Multifaceted Constraints

NeurIPS 2025 · Large language models › Agents, code and math · approach: Reasoning benchmarks and platforms · 0 citations (OpenAlex)

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 MAoP, a strategist-planner framework for wide-horizon thinking over multiple aspects in travel planning, plus Travel-Sim, an agent-based simulation benchmark that scores plans via feasibility and traveler-experience feedback.

Problem

LLM planning under many parallel, interacting real-world constraints (e.g., travel planning) where long-horizon step-by-step reasoning underperforms, and where static constraint pass-rate benchmarks ignore causal dependencies between itinerary steps.

Approach

Replace linear task decomposition with wide-horizon thinking: a strategist decomposes a request into multiple aspects, produces per-aspect guidance, and routes them into a planning blueprint that a planner executes over multi-turn aspect-focused dialogue; evaluation uses an LLM traveler agent executing plans in an event-driven sandbox, and the blueprint is distilled into one-step models.

Key results

• On Travel-Sim, the R1-Distill 7B strategist + R1-Distill 7B planner MAoP combination reaches FEA 60.7 and PER 81.4, versus 56.3 PER for the Qwen 2.5-7B + Qwen 2.5-32B combination.
• The distilled R1-Distill Qwen-7B student achieves FEA 73.7 and PER 84.2 with one-step wide-horizon thinking, surpassing the MAoP combinations of the same models.
• In preliminary experiments, DeepSeek-R1 with artificial guidance and wide-horizon thinking improves FEA from 52.6 (CoT) to 58.9 and PER from 62.5 to 68.0.

Measured effects reported in the paper

Surprising findings

• Qwen 2.5-7B strategist shows limited inference-time scalability, while the thinking-mode R1-Distill 7B strategist scales better with more aspects.
• Distilled 3B students (Qwen 2.5-3B) outperform MAoP combinations that use larger models, indicating distillation can exceed the multi-turn pipeline.

Limitations

• MAoP requires two models and multi-turn dialogue, costing more inference time than CoT; distillation is proposed to mitigate this.
• Weak strategists (under 7B) cause severe performance degradation.
• LLM traveler agent may be biased and unreproducible; Travel-Sim does not model force majeure or human-error emergencies.

Experimental setup

Datasets: Synthetic travel planning requests from Gemini-2.0-Pro-Exp-0205 and Qwen 2.5-32B plans Red Note (Xiaohongshu) travel blog posts Amap map API POI data

Benchmarks: Travel-Sim

Metrics: CPH (comprehensiveness) CPL (completeness) FEA (feasibility / TPSS) PER (personalization)

Baselines: Zero-shot CoT Plan-and-Solve Naive wide-horizon thinking RL with Long/Artifact guidance DeepSeek-R1 Gemini-2.5-Pro-Exp-03-25 TravelPlanner UnsatChristmas TravelAgent ITINERA ChinaTravel

Builds on: Chain-of-thought Plan-and-Solve GRPO ReAct ITINERA

Closest papers by meaning

• ATLAS: Constraints-Aware Multi-Agent Collaboration for Real-World Travel Planning (ICLR 2026 · similarity 0.90)
• WorldTravel: A Realistic Multimodal Travel-Planning Benchmark with Tightly Coupled Constraints (ICML 2026 · similarity 0.89)
• HiMAP-Travel: Hierarchical Multi-Agent Planning for Long-Horizon Constrained Travel (ICML 2026 · similarity 0.88)
• TravelPlanner: A Benchmark for Real-World Planning with Language Agents (ICML 2024 · similarity 0.87)
• SPIRAL: Symbolic LLM Planning via Grounded and Reflective Search (AAAI 2026 · similarity 0.86)
• Planning Anything with Rigor: General-Purpose Zero-Shot Planning with LLM-based Formalized Programming (ICLR 2025 · similarity 0.86)