[AAAI 2026 Oral] MOTIF: Multi-strategy Optimization via Turn-based Interactive Framework

---

📰 1. News

---

🔍 2. Overview

MOTIF is a turn-based, multi-agent framework for improving combinatorial optimization solvers by jointly evolving multiple algorithmic strategies rather than tuning a single heuristic.

Key Features

• 🎮 Competitive-Collaborative Learning: Two LLM agents take alternating turns to refine components
• 📊 Dynamic Baselines: Performance guided by adaptive baseline comparisons
• 🔄 Opponent Feedback: Each agent learns from the other's improvements
• 🧩 Structured Operators: Modular prompts enable targeted strategy refinement
• 🌐 Broad Search Space: Competitive dynamics encourage diverse adaptations

---

🚀 3. Quick Start

Step 1. Set Up Environment Variables

Create a .env file in the project root:

OPENAI_API_KEY="your-openai-api-key-here"

Step 2. Install Dependencies

pip install -r requirements.txt

Step 3. Run Default Experiment

python main.py

Step 4. Run Custom Experiments

Specify a solver:

python main.py solver=tsp_aco
python main.py solver=cvrp_dr_f1_f2_f3

Configure MCTS parameters:

python main.py \
       solver=tsp_aco \
       mcts.outer_iterations=20 \
       mcts.inner_iterations=10 \
       mcts.final_iterations=10

Customize LLM settings:

python main.py \
       llm.model=gpt-4o \
       llm.temperature=0.8

Step 5. Results

Results are saved to ./results/:

• F*_final_best.py: Final optimized strategy implementations
• *_round_*.json: Detailed experiment logs with performance metrics

---

📋 4. Supported Problems and Solvers

---

🔧 5. Customization and Extension

MOTIF is designed for easy extension. Follow these guides to add new problems and solvers.

Step 1. Create the problem directory:

problems/
└── problem_solver/
    ├── __init__.py
	├── solver.py        # Solver implementation
	├── eval.py          # Evaluation functions
	├── generator.py     # Instance generator
	├── prompts.py       # LLM prompts for each strategy
	├── F1.py            # Strategy component 1 (baseline)
	├── F2.py            # Strategy component 2 (baseline)
	├── F3.py            # Strategy component 3 (baseline)
	└── datasets/        # Training and test datasets

Step 2. Implement core files:

• prompts.py: Define PROBLEM_DESCRIPTION, CONSTRAINTS, and strategy prompts (F1, F2, F3)
• F*.py: Provide baseline implementations for each strategy component

Step 3. Create solver configuration:

# @package _global_.solver
base_path: ${paths.problems_dir}/problem_solver # Path to problem-solver directory

functions: [F1, F2] # Strategy files to optimize

---

📚 6. Citation

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

@inproceedings{nguyen2026motif,
  title     = {MOTIF: Multi-Strategy Optimization via Turn-based Interactive Framework},
  author    = {Kiet, Nguyen Viet Tuan and Tung, Dao Van and Dao, Tran Cong and Binh, Huynh Thi Thanh},
  booktitle = {Proceedings of the AAAI Conference on Artificial Intelligence (AAAI)},
  year      = {2026},
  address   = {Singapore},
  month     = {January},
  note      = {Oral Presentation}
}

---