LLM Colosseum: Language Models Play Street Fighter III

RESUMEN

En 30 segundos

Matthew Berman introduces LLM Colosseum, a project that connects language models to Street Fighter III and uses their chosen actions to control characters. The displayed match ends with a Mistral win; Berman says he has seen that result repeatedly in his own viewing. This is an illustrative game, not evidence that one model is generally superior or that its weights learned during the match. The video does not provide enough information to identify every model version or establish a statistical win rate.

1A fighting-game evaluationBerman presents Street Fighter III as an interactive way to compare language-model agents making decisions during a match.
2Observation to actionThe system forms a prompt from game information and a goal, asks an LLM for a move and applies the returned action in the game.
3A single displayed resultThe characters in the shown matchup are both Ken; the screen identifies OpenAI’s side and Mistral’s side. The victory screen reports Mistral won.

Conclusiones

LLM Colosseum uses game observations and prompts to request actions from language models.
The video presents one match between agents labeled OpenAI and Mistral, both using Ken.
The displayed result is a Mistral win, which does not by itself establish general model superiority.
The demonstration does not show that model weights were updated or that an agent learned during play.

TemasBenchmarks AI agents Experimentation RL environments & simulators

TEXTO COMPLETO

Lee el texto editado por capítulos

Ver transcripción

Solo la transcripción

A new game-based test

Berman introduces a Street Fighter III match between two language-model agents. He presents it as an alternative way to compare their decisions.

DIAPOSITIVA 1 / 2Game observations and action logThe display shows gameplay alongside the moves selected for each player.

I am watching two language-model agents play Street Fighter III, a game I love. The project, [LLM Colosseum](https://github.com/OpenGenerativeAI/llm-colosseum), presents this as another way to compare models through decisions made during a game. A short match is an illustration of that setup, not a broad assessment of model quality.

How the loop works

Game observations and a goal are used to create a prompt. The model returns a move that the game executes.

The loop takes observations from the game and a goal, constructs a prompt and asks the language model for the next move. The selected action is then applied in the game. That describes an agent using a model to choose actions; the video does not show the model’s weights being trained or updated during play.

The match result

Both sides use Ken in the demonstrated match, which ends with a Mistral victory. Berman hopes to see more characters and games.

DIAPOSITIVA 2 / 2OpenAI and Mistral matchThe game screen labels the opposing sides in the demonstrated Ken-versus-Ken match.

In the displayed matchup, I think the green Ken represents Mistral and the red Ken represents OpenAI; the game screen labels the sides. Both characters are Ken in this demo, and I hope the project adds more characters and games. The screen does not establish the exact version of the Mistral model.

After a brief pause at the end, the victory screen says Mistral won. I have seen this result repeatedly in my own viewing, but the short does not give a match count, controlled conditions or a statistical win rate. It is a fun demonstration, not proof that Mistral wins in general.