Track genomics, neuroscience and molecular biology.
Search papers, labs, and topics across Lattice.
2
0
5
LVLM judges, despite excelling in English, exhibit surprisingly inconsistent and unreliable behavior when evaluating content in other languages, revealing a critical blind spot in current alignment and evaluation pipelines.
Even explicit non-American prompting fails to dislodge the American English default that is structurally hardcoded across pretraining data, tokenizers, and prediction costs in modern LLMs.