Microsoft just released VibeVoice-ASR on Hugging Face A unified speech-to-text model that transcribes hour-long audio in one pass With built-in speaker diarization, timestamps, and customizable user context

Multiplex Thinking A branch-and-merge approach from Microsoft Research and UPenn. Samples K tokens per step into one multiplex token—self-adaptive: confident steps act like CoT, uncertain ones represent multiple paths. Beats discrete CoT with shorter sequences.

RubricHub A large-scale rubric dataset with ~110k instances for training open-ended generation models. Uses an automated coarse-to-fine framework to create highly discriminative evaluation criteria, enabling Qwen3-14B to surpass GPT-5 on HealthBench.