1 paper
Linjuan Wu, Haoran Wei, Jialong Tang +4
As LLMs reduce English-centric bias, a surprising trend emerges: non-English responses sometimes outperform English on reasoning tasks. We hypothesize that language functions as a…