Learning from Synthetic Laboratory: Language Models as Auction Participants

Table of Contents

1. Overview
2. Getting Started
3. Auction Parameters
4. Reproducing Our Results

---

Overview

This repository accompanies our paper → https://openreview.net/forum?id=XZ71GHf8aB.

If you build on this work, please cite us:

@article{zhuevidence,
  title  = {Evidence from the Synthetic Laboratory: Language Models as Auction Participants},
  author = {Zhu, Kehang and Shah, Anand V and Jiang, Yanchen and Horton, John Joseph and Parkes, David C}
}

---

Getting Started

Our code supports Python 3.9 – 3.11.

# 1  Create and activate a virtual‑env
python -m venv venv
source venv/bin/activate

# 2  Install dependencies
pip install edsl

# 3  Configure your OpenAI key
cat > .env <<'EOF'
OPENAI_API_KEY=YOUR_KEY_HERE
EOF

---

Auction Parameters

All experiments are launched via one of the three driver scripts below. Use the flags shown to configure auction mechanics.

Auction type	Driver script	Required flags	Optional flags
Sealed‑bid	main.py	--seal_clock seal	--price_order {first,second,third,allpay} --private_value {private,affiliated,common}
Clock	main.py	--seal_clock clock --ascend_descend descend	--open_blind {open,blind} --private_value {private,affiliated,common}
Ebay proxy	main_ebay.py	--seal_clock ebay (fixed) --price_order second (fixed) --private_value private (fixed)	--turns 10 --closing {true,false} --reserve_price 60
Intervention study	main_intervention.py	(inherits flags from sealed‑bid)

Quick examples

Run a second‑price sealed‑bid auction:

python main.py 

Run a second‑price sealed‑bid auction with intervention:

python main_intervention.py 

Run the Ebay‑style proxy auction:

python main_ebay.py

You can vary all the hyperparameters in these main functions.

Reproducing Our Results

We use EDSL, whose universal remote cache stores every completed LLM call. Re‑running our code with the same prompt and same random seed therefore incurs no additional API cost—results are retrieved automatically.

1  Seeds and model settings

Value regime	Seeds used
Private	1299 – 1309
Affiliated & Common	1399 – 1409

All experiments use GPT‑4, temperature = 0.5.

2  Cached runs (default)

results = survey.by(model).run(
    remote_inference_description="cache reuse", 
    remote_inference_visibility="public"         
)

The snippet above will first look in the cache; if a match is found, the result is loaded instantly.

3  Forcing a fresh run (optional)

To ignore the cache—for instance, when testing a new prompt—add fresh=True:

results = survey.by(model).run(
    remote_inference_description="fresh run",
    remote_inference_visibility="public",
    fresh=True
)

4  Verifying cache hits

EDSL prints the Job UUID and whether it was served from cache. You can also inspect the universal cache via the web UI linked in the EDSL docs.

---

Happy experimenting! If anything is unclear, please open an issue 🙌

🔧 Dependencies

The main third-party package requirement are openai and edsl.

💡 Contributing, Feature Asks, and Bugs

Interested collaborating in LLM as auction participants? Found a nasty bug that you would like us to squash? Please send us an email at kehangzhu@gmail.com.