SAGE

SAGE 🚀🚀🚀

An LLM agent that plays imperfect-information games which captures opponents’ behavior tendencies and react over them!

supported games ✅ Leduc Hold’em ✅ Liar’s Dice ✅ Goofspiel

Quick Start

Prepare the enviroment

git clone
cd SAGE
conda create -n SAGE python=3.11 -y
conda activate SAGE
pip install -r requirements.txt

optionally, you can test whether the code runs correctly:

pytest tests/ -q

API preparation:

export OPENAI_API_KEY="<your key>"
export OPENAI_BASE_URL="<base url>"

prepare the CFR model

# Leduc Hold'em
python runs/leduc/train_cfr.py --help
python runs/leduc/train_cfr.py --output-dir results/leduc/cfr_model

# Small Liar's Dice setting
python runs/liars_dice/train_cfr_2x3.py --help
python runs/liars_dice/train_cfr_2x3.py

# Goofspiel MCCFR
python runs/goofspiel/openspiel_mccfr_goofspiel.py --help

Option 1 : Human against Agent!

If you want to play against the LLM instead of letting’em self playing, you can try:

python main.py leduc \
  --p1 Human --p1-method human \
  --p2 CFR --p2-method cfr \
  --seed 42

Use --verbose if you want to inspect the trajectory after the game.

Option 2: Agents battlefield To run the full benchmark, you can refer to the scripts below:

python runs/leduc/leduc_tom_ablation.py --help
bash ablation/run_leduc_vs_baselines.sh --dry-run

For lightweight benchmark summaries, use stats/benchmark_summary.py. It reads completed run summary.json files and prints performance rows; the full research analysis and plotting pipeline is not included in the public repository.

Each recorded run should have this shape:

results/<game>/<experiment>/<run_id>/
├── config.json
├── summary.json
├── trajectories/
├── metrics/
├── figures/
├── logs/
└── prompts/

A quick look at the comparison 📊

The table below is the selected Leduc comparison reported in the manuscript evaluation. It shows average P0 payoff per game in the 180-game evaluation setting; it is included as a navigation aid, not as a replacement for the full experimental record.

Agent condition	Avg. P0 payoff/game
Vanilla LLM	1.56
LLM + equilibrium guidance	1.91
LLM + opponent modeling	1.86
Suspicion baseline	1.43
SAGE	2.28