🕷️
SetupAI
📄較早收集於 13h

TrajGenAgent：用於合成移動數據的階層式 LLM Agent

TrajGenAgent：用於合成移動數據的階層式 LLM Agent
PostLinkedIn
📄閱讀原文： ArXiv AI
#synthetic-data#human-mobility#llm-agents#urban-planningtrajgenagenttrajgenagentllm

💡利用這個新的階層式 LLM Agent 框架，無需昂貴的微調即可生成逼真的人類移動數據。

⚡ 30-Second TL;DR

有什麼變化

採用兩階段的協調者-執行者架構來進行軌跡生成。

為什麼重要

這項研究為生成合成移動數據提供了一種具成本效益的替代方案，對於隱私與數據稀缺為主要障礙的城市規劃與交通研究至關重要。

下一步行動

如果您正在進行城市模擬或合成數據生成，請評估 TrajGenAgent 的兩階段架構作為基準，以避免昂貴的微調週期。

誰應關注：Researchers & Academics

關鍵要點

  • •採用兩階段的協調者-執行者架構來進行軌跡生成。
  • •利用上下文學習（In-context learning）避免微調帶來的計算成本。
  • •整合了用於 POI 檢索與旅行時間傳播的確定性工作流程。
  • •引入異常檢測框架來評估行為與語義的合理性。

🧠 深度解析

Web-grounded analysis with 17 cited sources.

🔑 增強重點摘要

  • •The TrajGenAgent framework, formally termed "Narrative-to-Action," employs a multi-layer cognitive process that integrates high-level narrative reasoning, mid-level reflective planning, and low-level behavioral execution to generate synthetic mobility data.
  • •It utilizes a 'creative writer' agent to produce rich, diary-style narratives of motivations and contexts, which are then converted into machine-readable plans by a 'structural parser' agent.
  • •The framework incorporates a dynamic execution module that grounds agents in geographic environments and facilitates adaptive behavioral adjustments using a novel metric called Mobility Entropy by Occupation (MEO), which accounts for varying schedule flexibility across different occupations.
  • •This approach represents a shift from traditional data-driven synthetic mobility generation to a cognition-driven simulation, providing not only realistic trajectories but also interpretable representations of human decision logic.

🛠️ 技術深入

  • Architecture: Hierarchical LLM-Agent Framework, termed "Narrative-to-Action," designed to mimic human cognitive processes for travel decisions.
  • Macro-Level Agents:
    • Creative Writer Agent: Generates high-level, diary-style narratives detailing motivations and contexts for mobility.
    • Structural Parser Agent: Translates these narratives into structured, machine-readable plans.
  • Micro-Level Execution:
    • Dynamic Execution Module: Grounds agents within simulated geographic environments.
    • Adaptive Behavioral Adjustments: Guided by the Mobility Entropy by Occupation (MEO) metric, which quantifies heterogeneous schedule flexibility based on occupational personalities.
    • Action Execution: Agents perform concrete actions such as selecting locations, transportation modes, and time intervals through interaction with an environmental simulation.
  • Core Principle: Integrates high-level narrative reasoning, mid-level reflective planning, and low-level behavioral execution within a unified cognitive hierarchy.
  • Learning Paradigm: Leverages in-context learning to avoid the computational costs associated with model fine-tuning.

🔮 前景展望AI analysis grounded in cited sources

The TrajGenAgent framework could significantly enhance urban planning and traffic management simulations.
By generating realistic and interpretable human mobility patterns, it offers scalable insights for societal problems such as urban planning and traffic management.
This cognition-driven approach will lead to more robust and adaptable synthetic mobility data for diverse applications.
Moving from a data-driven to a cognition-driven simulation provides a scalable pathway for understanding, predicting, and synthesizing complex urban mobility behaviors.

⏳ 時間線

2022-04
Early research on generating synthetic mobility data using deep recurrent neural networks (RNNs) is published, addressing utility and privacy concerns.
2024-01
STG-LLM is introduced, an approach empowering LLMs for spatial-temporal forecasting by tokenizing graph data.
2025-02
TrajLLM, a modular LLM-enhanced agent-based framework for realistic human trajectory simulation, is published.
2025-10
The paper "From Narrative to Action: A Hierarchical LLM-Agent Framework for Human Mobility Generation" (TrajGenAgent) is published on ArXiv.
2025-10
TrajAgent, an LLM-Agent Framework for Trajectory Modeling via Large-and-Small Model Collaboration, is published on ArXiv and accepted for NeurIPS 2025.
2026-03
A comprehensive survey titled "LLMs for Human Mobility: Opportunities, Challenges, and Future Directions" is published, summarizing the rapidly evolving field.

📎 來源 (17)

Factual claims are grounded in the sources below. Forward-looking analysis is AI-generated interpretation.

  1. arxiv.org
  2. arxiv.org
  3. researchgate.net
  4. emergentmind.com
  5. upenn.edu
  6. aclanthology.org
  7. arxiv.org
  8. arxiv.org
  9. mit.edu
  10. arxiv.org
  11. arxiv.org
  12. github.com
  13. neurips.cc
  14. arxiv.org
  15. deeplearn.org
  16. arxiv.org
  17. arxiv.org
📄閱讀原文： ArXiv AI
📰

AI 週報

閱讀本週精選 AI 大事摘要 →

👉相關動態

同主題

探索 #synthetic-data

同產品

更多關於 trajgenagent

同來源

最新來自 ArXiv AI

AI 策展新聞聚合。所有內容版權歸原始發布者所有。
原始來源： ArXiv AI ↗