Google researchers ship Hebatron, Vertex-Softmax verification, and well-being-from-speech LLM study

Google research dropped three papers today: Hebatron, a Hebrew-specialized open-weight MoE LLM built on NVIDIA Nemotron-3 with a two-million-sample bilingual SFT set; Vertex-Softmax, a tighter certified-verification method for Transformer attention; and a zero-shot LLM study predicting Ryff well-being scores from a few minutes of speech.