GPlan: Generative Spatiotemporal Intent Sequence Recommendation via Implicit Reasoning in Amap

AMAP, Alibaba Group

📋 Overview

Progressive Implicit CoT Distillation (PICD) training framework of GPlan. It uses curriculum learning to compress structured CoT text into fixed-length latent think-token blocks and uses a compression-aware learning-rate schedule (CALR) for the structure-to-polish transition.

📊GSISR Dataset

The GSISR dataset is collected from Amap and provided in data_process/dataset/. All data have been anonymized to protect privacy — original feature names, POI identifiers, and user identifiers have been replaced with generic placeholders.

File	Description	Num
data_process/dataset/train.csv	Training set	100,000
data_process/dataset/test.csv	Test set	1,000

User Profiles and Behavior History

Each user is described by 14 anonymized profile features. All categorical values have been mapped to numerical IDs.

Field	Description
User ID	A unique numerical identifier for each user.
Profile Feature 1–14	Anonymized user profile attributes.
Short-term Behavior Seq	Anonymized short-term behavior sequence. POI names and behavior types (e.g., click) are replaced with numerical IDs (p_, act_).
Long-term Behavior	Anonymized long-term behavior feature. Original values are replaced with numerical IDs.

Context Information

Field	Description
Current Time	Time of the request.
Weekend Flag	Whether the current day is a weekend (0/1).
Holiday Flag	Whether the current day is a holiday (0/1).
Current City & District	The city and district where the user is located.
Current POI Name	The name of the user's current Point of Interest (mapped to ID, p_).
Current POI Category	The tag of the current POI.

Trigger Events

Each request includes 7 trigger event features that capture the user's immediate intent signals.

Field	Description
Trigger 1–7	Anonymized event trigger features. Original event types and descriptions are replaced with numerical IDs or kept as timestamps.

Labels

Each label is an intent sequence — a JSON array of tool-calling intents representing the recommendation. Each intent includes a tool name and associated parameters selected from a predefined intent library:

[
  {"工具名称": "tool_5", "起始位置": "当前位置", "空间范围": "附近", "tag": "美食"},
  {"工具名称": "tool_2", "起始位置": "当前位置", "终点位置": "家"},
  ...
]

The intent library includes 10 tool types covering scenarios such as ride-hailing, navigation, transit, POI recommendation, order reminders, weather queries, etc.

PICD Training Data Preparation

The released CSV files provide final intent sequences as labels. To run PICD training, prepare a CoT-augmented CSV with the same schema as train.csv, where raw_labels contains a structured CoT followed by the final intent sequence.

Step 1: Prepare structured CoT

For each training sample, generate a concise reasoning trace from the user profile, behavior history, current context, and gold intent sequence. The CoT should explain why the plan is reasonable, with each <STEP_n> aligned to the n-th intent in the JSON label:

<THOUGHT>
<CONTEXT>Briefly analyze the current context and user profile</CONTEXT>
<STRATEGY>Describe the planning strategy</STRATEGY>
<STEP_1>Explain the first recommended intent</STEP_1>
...
<STEP_n>Explain the n-th recommended intent</STEP_n>
</THOUGHT>

The number of <STEP_n> fields should match the number of intents in the JSON array.

Step 2: Write raw_labels

Concatenate the CoT and JSON intent sequence in raw_labels:

<THOUGHT><CONTEXT>...</CONTEXT><STRATEGY>...</STRATEGY><STEP_1>...</STEP_1><STEP_2>...</STEP_2><STEP_3>...</STEP_3></THOUGHT>[{"工具名称":"tool_5","起始位置":"当前位置","空间范围":"附近","tag":"美食"},{"工具名称":"tool_2","起始位置":"当前位置","终点位置":"家"},{"工具名称":"tool_7","tag":"景点"}]

The collator parses this field and applies progressive implicit CoT distillation automatically.

🚀 Quick Start

Evaluation on the Public Dataset

pip install -r requirements.txt
bash test.sh

The test script reports the offline metrics used in the paper: Acc@1, NDCG@3, and NES (normalized edit similarity). It can use the JSON-only labels included in data_process/dataset/test.csv.

PICD Training

TRAIN_CSV=/path/to/cot_augmented_train.csv bash finetune.sh

finetune.sh uses --cot_mode=latent_multi_cot and expects CoT-augmented raw_labels.

📁 Project Structure

├── finetune.py                         # Training script (WeightedLossTrainer + SyncEpochCallback)
├── finetune.sh                         # Training launch script
├── test.py                             # Test script (Acc@1, NDCG@3, NES)
├── test.sh                             # Test launch script
├── data_process/
│   ├── dataset/
│   │   ├── train.csv                   # Training dataset (anonymized)
│   │   └── test.csv                    # Test dataset (anonymized)
│   ├── collate_fns.py                  # PICD data collator
│   └── data_loader.py                  # CSV data loading
├── utils.py                            # Utility functions and argument definitions
├── add_tokens/extended_cot_vocabs.json  # CoT special token vocabulary
├── config/ds_z3_bf16.json              # DeepSpeed ZeRO-3 configuration
└── requirements.txt

📝 Citation

If you find our work useful in your research, please consider citing:

@misc{wang2026generative,
    title={Generative Spatiotemporal Intent Sequence Recommendation via Implicit Reasoning in Amap},
    author={Sicong Wang and Ruiting Dong and Yue Liu and Bowen Zheng and Jun Meng and Jie Li and Shuaijun Guo and Yu Gu and Fanyi Di and Xin Li},
    year={2026},
    eprint={2605.28888},
    archivePrefix={arXiv},
    primaryClass={cs.IR}
}