Game Arena: Strategic LLM Evaluation in Competitive Environments
Kaggle’s Game Arena tests LLMs by making them compete in games where the benchmark can keep getting harder.
The technical report describes an open evaluation platform built around head-to-head play rather than fixed task sets. Its first environments are Chess, Poker, and Werewolf, covering perfect information, hidden information, and multiplayer deception settings. The authors frame the setup as a way to measure strategic planning, adaptation, and robustness under uncertainty without saturating as quickly as static benchmarks. ArXiv · AI/CL/LG's note
The technical report describes an open evaluation platform built around head-to-head play rather than fixed task sets. Its first environments are Chess, Poker, and Werewolf, covering perfect information, hidden information, and multiplayer deception settings. The authors frame the setup as a way to measure strategic planning, adaptation, and robustness under uncertainty without saturating as quickly as static benchmarks. ArXiv · AI/CL/LG's note
score 6