AMAP Agentic Planning Technical Report

TL;DR

The paper introduces STAgent, an agentic LLM designed for real-world spatio-temporal reasoning that can interact with a large toolset. It presents a three-part framework: a stable, scalable tool environment; a hierarchical, seed-driven data curation pipeline with a dynamic difficulty-aware curriculum; and a cascaded training recipe combining supervised fine-tuning with RL (SFT-Guided RL) built on Qwen3-30B-A3B. The approach yields strong performance on TravelBench and maintains generalization across diverse benchmarks, validating the effectiveness of integrating tool invocation, data-aware curriculum design, and RL to push agentic capabilities in open-world scenarios. This work contributes a practical paradigm for developing specialized agents in real-world domains and provides insights into maintaining general-purpose abilities while optimizing domain-specific skills.

Abstract

We present STAgent, an agentic large language model tailored for spatio-temporal understanding, designed to solve complex tasks such as constrained point-of-interest discovery and itinerary planning. STAgent is a specialized model capable of interacting with ten distinct tools within spatio-temporal scenarios, enabling it to explore, verify, and refine intermediate steps during complex reasoning. Notably, STAgent effectively preserves its general capabilities. We empower STAgent with these capabilities through three key contributions: (1) a stable tool environment that supports over ten domain-specific tools, enabling asynchronous rollout and training; (2) a hierarchical data curation framework that identifies high-quality data like a needle in a haystack, curating high-quality queries with a filter ratio of 1:10,000, emphasizing both diversity and difficulty; and (3) a cascaded training recipe that starts with a seed SFT stage acting as a guardian to measure query difficulty, followed by a second SFT stage fine-tuned on queries with high certainty, and an ultimate RL stage that leverages data of low certainty. Initialized with Qwen3-30B-A3B to establish a strong SFT foundation and leverage insights into sample difficulty, STAgent yields promising performance on TravelBench while maintaining its general capabilities across a wide range of general benchmarks, thereby demonstrating the effectiveness of our proposed agentic model.

Figures (7)

• Figure 1: The overall framework of STAgent. It presents a comprehensive pipeline designed for real-world spatio-temporal reasoning. The framework consists of three key phases: (1) Robust Interactive Environment, supported by the ROLL infrastructure and FastMCP protocol to enable efficient, asynchronous tool-integrated reasoning. (2) High-Quality Data Construction, which utilizes a self-evolving selection framework to filter diverse and challenging queries from massive unsupervised data; (3) Cascade Training Recipe, an SFT-Guided RL paradigm that categorizes samples by difficulty to synergize supervised fine-tuning with reinforcement learning.
• Figure 2: Typical reasoning tasks under spatio-temporal scenario.
• Figure 3: A visual taxonomy of our intent classification system. The hierarchical structure is organized into five primary categories: Rules and Policies, Discovery, Planning and Decision, Dynamic Information, and Application Interaction. The taxonomy further branches into 16 second-level categories and terminates in 30 fine-grained leaf nodes, capturing the multi-faceted complexity of real-world user queries in navigation and travel scenarios.
• Figure 4: Fine-grained Difficulty Distribution across 30 Geospatial Domains. The visualization reveals the distinct complexity profiles of different tasks. While atomic queries (e.g., basic attributes) cluster in low-difficulty regions (Score 1-2), composite tasks (e.g., long trip itinerary) exhibit a significant proportion of high-complexity reasoning (Score 4-5). The visible segments of Score -1 and 0 (denoted as Ext.1) across domains like System Control highlight our active sampling of boundary cases for hallucination mitigation.
• Figure 5: The training procedure in the SFT phase.