Multilingual Self-Play Quantifies Cross-Lingual Skill Inconsistency in LLMs
August 25, 2026
A multilingual self-play framework using TextArena isolates language effects on model behavior by having identical models compete in text-based games via different language interfaces. The study evaluates how spatial reasoning and resource allocation skills fluctuate across eight languages for open-weight models.
HOW THIS AFFECTS YOU
●
researcherYou can use this framework to decouple language-specific knowledge from core reasoning capabilities.