2026-09-28 · America/Los_Angeles · 论文 · #19

Game Arena: Strategic LLM Evaluation in Competitive Environments

We introduce Kaggle Game Arena, an open and ever-expanding platform to evaluate large language models (LLMs) through competitive games. Different from static benchmarks, game arena enables models to play head-to-head matchups in structured environments where the gameplay strength naturally increases as models evolve, preventing performance saturation. This technical report details the infrastructure behind Game Arena and describes the three pilot game environments: Chess, Poker, and Werewolf. These environments span perfect information, imperfect information, and multiplayer game settings,…

热度 35.2 / 100；排名与评分保留该期记录。