Search papers, labs, and topics across Lattice.
Johns Hopkins University
3
0
3
4
MT models show a troubling bias towards US content, with specialized models proving less robust across diverse locales.
Retaining the original question in multilingual reasoning cascades can dramatically enhance performance, revealing that context is key to effective translation and reasoning.
Forget scaling laws: a specialized 8B parameter translation model can outperform a 70B general-purpose LLM on 1,600 languages.
