Matthew Berman Tries LLM Colosseum, a Street Fighter Model Benchmark

RESUMEN

En 30 segundos

Matthew Berman introduces LLM Colosseum, an open-source project that uses a fighting-game emulator to pit language models against each other. He reviews the project’s historical ranking, speculates that response speed matters and demonstrates the installation and one match. The displayed results are specific to this real-time game harness and do not establish a general ranking of model capability.

1A real-time matchBerman shows Mistral Small against Mistral Medium and speculates that faster responses can matter.
2A historical leaderboardThe snapshot ranks GPT-3.5 Turbo above the other displayed models for this game, not for all tasks.
3Game state to next moveThe harness turns observations and possible actions into a text prompt, then executes the model’s chosen move.
4Setting up the emulatorThe walkthrough uses DIAMBRA Arena, Docker and a Python environment before configuring model APIs.
5One demonstration resultBerman reports a Mistral win in the displayed match; one game does not establish broad model superiority.

Conclusiones

LLM Colosseum sends text descriptions of game state to models and executes their selected moves.
The displayed 2024 leaderboard is specific to the project’s fighting-game evaluation.
Berman suspects inference speed affects outcomes, but does not isolate it experimentally.
The walkthrough uses DIAMBRA Arena, Docker, Python, a compatible ROM and model API credentials.
A single demo match does not determine general model quality.

TemasEvaluaciones Multi-agent systems Experimentation Inference

TEXTO COMPLETO

Lee el texto editado por capítulos

Ver transcripción

Solo la transcripción

Models fight in Street Fighter III

Berman introduces LLM Colosseum, a match and the historical ranking. He speculates about the importance of response speed.

LLM Colosseum pits language models against each other in Street Fighter III, with each model choosing moves during the game. Stan Girard and the Quivr team created the open-source project. It is an entertaining alternative way to compare model behavior, and I will also show the setup used in April 2024.

DIAPOSITIVA 1 / 6Two model-controlled charactersThe opening example shows Mistral Small and Mistral Medium controlling Ken characters.

In the opening match, I identify red Ken as Mistral Small and green Ken as Mistral Medium; Small wins. I suspect response speed plays a major role in a real-time fighting game and may help explain why smaller models sometimes win. That is my interpretation, not an isolated measurement proving speed caused the result.

The project presents Elo rankings and a head-to-head win-rate matrix. In the snapshot shown, GPT-3.5 Turbo leads several GPT-4 and Mistral models. I find that surprising, but I am not sure exactly what the benchmark measures; rankings in this particular game should not be treated as a general ranking of language-model intelligence.

DIAPOSITIVA 2 / 6Historical Elo rankingThe displayed 2024 project snapshot ranks the tested models on this game.

How the game harness works

The project turns game observations and possible actions into text prompts for each model. The emulator then executes the selected moves.

The creators propose speed, strategic thinking, adaptability and resilience as relevant traits. The system turns game observations, available actions and goals into a prompt; each model receives a text description of the game state and selects the next move. A game emulator executes those actions in real time.

DIAPOSITIVA 3 / 6Head-to-head win matrixThe matrix compares observed win rates among the listed model pairings.

DIAPOSITIVA 4 / 6Game observation and action loopThe diagram shows game state and possible actions forming a prompt for the model’s next move.

Installation and a demo match

Berman walks through DIAMBRA, Docker, Python and API setup. He configures models and shows one match, whose outcome is specific to that run.

To set it up, I clone the LLM Colosseum repository in VS Code and install the DIAMBRA Arena emulator. I register for its service, install Docker and create a conda environment with Python 3.10. After activating the environment, I install DIAMBRA Arena and list supported ROMs to check the installation.

The tutorial requires a compatible game ROM placed as a ZIP file in the emulator’s ROMs folder. I do not provide a download location and discuss game ownership. Rights to obtain and use a ROM depend on the applicable license and law; owning a physical copy alone does not establish permission to download any copy.

I install the LLM Colosseum requirements, reopen the project and reactivate the conda environment. I rename the example environment file to .env and configure API credentials for OpenAI and Mistral. The recording’s keys are described as being revoked before publication; the tutorial’s service interfaces and model names reflect 2024.

DIAPOSITIVA 5 / 6Configuring the playersThe demo code contains model settings for the two players.

In demo.py I configure one player to use a Mistral model and the other GPT-3.5 Turbo, then run the demo. A first run may require a DIAMBRA login, and the emulator takes some time to start. In the match shown, I initially mix up the player colors before correcting myself.

DIAPOSITIVA 6 / 6A local demo matchThe emulator displays a match between the configured models.

The models choose moves as the two Ken characters fight. I say Mistral wins this particular match and note that the project at the time offers only Ken in this setup. One match and the displayed leaderboard are examples of this game harness, not controlled evidence that a model is better at other tasks.