Multilingual Self-Play Reveals Cross-Lingual Skill Inconsistencies in LLMs

The-CoLab · hf · 2026-08-27

The study "Skill Issue" uses multilingual self-play to demonstrate that Large Language Models exhibit significant cross-lingual skill inconsistencies in reasoning and strategy. It finds that models perform differently depending on the language used. Notably, these inconsistencies can be partially recovered by altering the intermediate reasoning language.

Original post →

More from Research

Research channel →