MAGIC-HMO

Heterogeneous Multi-Objective Optimization for Short-form Creative NLG

🧠 Teaser

MAGIC-HMO is a training-free multi-agent framework that formulates short-form creative generation as heterogeneous multi-objective optimization, jointly optimizing:

• 🎯 generation quality
• 📌 user constraints
• 🔍 explanation reliability

🔥 Motivation

Real-world creative generation (e.g., Chinese naming) requires balancing:

• semantic meaning
• cultural expectations
• phonetics & structure
• interpretability

⚠️ Existing approaches:

• optimize single objective
• ignore reliable explanations

➡️ We propose multi-objective optimization with explanation awareness

✨ Contributions

• New formulation: CNLG → Heterogeneous Multi-Objective Optimization (HMO)

• Dual-objective modeling:

  • Explicit User Objectives (EUOs)
  • Implicit Reliability Objectives (IROs)

• Training-free multi-agent framework

• Explanation-aware iterative optimization

• Compatible with multiple LLM backbones

🏗 Framework

MAGIC-HMO consists of three collaborative agents:

🧩 MOM — Multi-Objective Manager

• Query understanding
• Knowledge retrieval (e.g., poetry)
• Objective refinement

✍️ MOG — Multi-Objective Generator

• Generates:

  • result
  • structured explanation

🧪 MOE — Multi-Objective Evaluator

• Evaluates:

  • IROs first
  • EUOs second

• Provides feedback for iterative refinement

⚖️ Multi-Objective Perspective

MAGIC-HMO searches for a balanced solution:

• satisfies EUOs (user intent)
• ensures IROs (reliability)

➡️ Equivalent to finding a Pareto-optimal point in objective space

🔍 Case Study

• Demonstrates iterative optimization
• Shows how explanations guide refinement
• Achieves progressively better multi-objective satisfaction

🚀 Quick Start

Environment

conda create -n magic-hmo python=3.10 -y
conda activate magic-hmo

Dataset

The datasets used in this project are available on Hugging Face:

• Chinese Poems: https://huggingface.co/datasets/Audrey33/Chinese_poems
• Chinese Hanzi Poem: https://huggingface.co/datasets/Audrey33/Chinese_hanzi_poem

Due to size limitations, the full datasets are not included in this repository.

Run

python MagicHMO.py --backbone qwen --mode batch --number 20

Single Query

python MagicHMO.py \
  --backbone qwen \
  --mode single \
  --query "请为一个于2024年6月1日10:00出生的女孩起名..."

📂 Structure

MagicHMO.py
Agents.py
utils/
RetrivalPoems.py
data/
figs/

📌 Key Insight

MAGIC-HMO is not just generating text — it is searching for explainable and reliable solutions under multiple objectives

📚 Citation

@article{magic_hmo,
  title={MAGIC-HMO: Heterogeneous Multi-Objective Optimization for Short-form CNLG},
  author={Anonymous},
  year={2025}
}