Yinghan Hou

Accuracy metrics can mask critical failure states in language models, revealing that nearly 11% of outputs can terminate without a final answer under specific conditions.

Judge upgrades in LLMs aren't interchangeable—only specific parameter increases yield reliable improvements in evaluation consistency.