STORM: Official Repository for "WHEN TO ACT, WHEN TO WAIT"

Structured Task-Oriented Representation Model for Intent Triggerability
in Task-Oriented Dialogue Systems

Welcome to the official repository for STORM, a framework that models asymmetric information dynamics between user and agent LLMs to track and trigger user intents effectively in task-oriented dialogue.

       

---

Authors

Yaoyao Qian     Jindan Huang     Yuanli Wang     Simon Yu
Kyrie Zhixuan Zhou     Jiayuan Mao     Mingfu Liang     Hanhan Zhou

Northeastern University     Tufts University     Boston University     University of Texas at San Antonio
Massachusetts Institute of Technology     Northwestern University     George Washington University

📧  Corresponding author: Yaoyao Qian

---

Overview of the STORM Framework

---

Core Features

• Asymmetric dialogue generation using separate LLM models
• RAG (Retrieval-Augmented Generation) support
• Emotional state and intent tracking
• Inner thoughts generation
• Comprehensive dialogue analysis and reporting
• Batch processing capabilities
• Profile-based dialogue generation

Project Structure

storm/
├── src/                    # Core source code
├── scripts/               # Utility scripts
│   ├── setup.sh          # Environment setup script
│   ├── run_demo.sh       # Demo dialogue generation
│   ├── run_convert.sh    # Data conversion utilities
│   ├── run_profile_generate.sh    # Profile generation
│   ├── run_turn_analysis.sh      # Turn-by-turn analysis
│   ├── import_and_test_rag.sh    # RAG system setup and testing
│   └── run_batch_dialogues.sh    # Batch dialogue generation
├── example_data/         # Sample data and configurations
├── tests/               # Test suite
└── requirements.txt     # Python dependencies

If you want to run it, please move the example_data/profiles folder to the root directory first.

Quick Start

1. Clone the repository:

   cd storm

2. Run the setup script:

   ./scripts/setup.sh

3. Set your API keys:

   export OPENAI_API_KEY='your_openai_api_key'
   export OPENROUTER_API_KEY='your_openrouter_api_key'

Scripts Overview

Setup and Configuration

• setup.sh: Initializes the environment and installs dependencies
• run_profile_generate.sh: Generates user profiles for dialogue generation

Core Functionality

• run_demo.sh: Runs a single dialogue generation demo
• run_batch_dialogues.sh: Processes multiple dialogues in batch mode
• run_turn_analysis.sh: Analyzes dialogue turns and generates reports

RAG System

• import_and_test_rag.sh: Sets up and tests the RAG system
• run_convert.sh: Converts and processes dialogue data

Usage Examples

Running a Demo Dialogue

./scripts/run_demo.sh 

Batch Processing

./scripts/run_batch_dialogues.sh 

RAG System Setup

./scripts/import_and_test_rag.sh 

---

❤️ Acknowledgement

We gratefully acknowledge the 🐪 CAMEL codebase for providing essential resources that supported this work.

---

📚 Citation

If you use this work, please cite our paper:

@misc{qian2025actwaitmodelingstructural,
  title={WHEN TO ACT, WHEN TO WAIT: Modeling Structural Trajectories for Intent Triggerability in Task-Oriented Dialogue}, 
  author={Yaoyao Qian and Jindan Huang and Yuanli Wang and Simon Yu and Kyrie Zhixuan Zhou and Jiayuan Mao and Mingfu Liang and Hanhan Zhou},
  year={2025},
  eprint={2506.01881},
  archivePrefix={arXiv},
  primaryClass={cs.AI},
  url={https://arxiv.org/abs/2506.01881}
}

📬 Contact

For questions, bug reports, or collaboration inquiries, please contact:
Yaoyao(Freax) Qian — qian.ya@northeastern.edu