lechmazur/step_game

Multi-Agent Step Race Benchmark: Assessing LLM Collaboration and Deception Under Pressure. A multi-player “step-race” that challenges LLMs to engage in public conversation before secretly picking a move (1, 3, or 5 steps). Whenever two or more players choose the same number, all colliding players fail to advance.

27
/ 100
Experimental

This project offers a competitive "step-race" game where three AI models (Large Language Models) engage in public conversation before secretly choosing to advance 1, 3, or 5 steps. If multiple models pick the same number, they all fail to advance that turn. It helps AI researchers and developers understand how different LLMs strategize, collaborate, and deceive under pressure by observing their conversational tactics and resulting moves on the board.

Use this if you are an AI researcher or developer looking to evaluate the social reasoning, negotiation, and strategic capabilities of Large Language Models in a dynamic, multi-agent environment.

Not ideal if you are looking for a simple benchmark of factual recall or a tool to test traditional coding abilities of LLMs.

LLM evaluation multi-agent systems AI strategy games social reasoning AI deception
No License No Package No Dependents
Maintenance 6 / 25
Adoption 9 / 25
Maturity 8 / 25
Community 4 / 25

How are scores calculated?

Stars

85

Forks

2

Language

License

Last pushed

Dec 09, 2025

Commits (30d)

0

Get this data via API

curl "https://pt-edge.onrender.com/api/v1/quality/agents/lechmazur/step_game"

Open to everyone — 100 requests/day, no key needed. Get a free key for 1,000/day.