Cross-Lingual Consensus: Aligning Multilingual Cultural Knowledge via Multilingual Self-Consistency
Researchers have developed two distinct methods to improve the cultural awareness of large language models. One approach, used by DFKI-MLT for SemEval-2026 Task 7, employs activation steering with language vectors to adapt models at inference time, achieving 86.96% accuracy in the multiple-choice track. The other method, termed Cross-Lingual Consensus, uses multilingual self-consistency and self-critique to surface and propagate latent cultural knowledge from local-language representations to English prompts, boosting performance on the BLEnD benchmark by an average of 5.03%. Both studies highlight the challenge of uneven cultural knowledge in LLMs and propose novel techniques to address it. AI
IMPACT These methods could lead to more equitable and globally relevant AI systems by reducing Western-centric biases in LLMs.