Apple researchers present what they describe as the first masked diffusion models pretrained from scratch on text, image-text and audio-text data. The paper studies multimodal scaling laws, data mixing ratios, noise schedules and batch-size effects, and proposes default settings for inference.
Apple researchers propose a lightweight way to spot code-switching, when speakers alternate languages mid-sentence, directly from audio rather than from transcripts. The method relies on frame-level language signals inside pretrained speech models and aims to help gather more such examples for training data.
Latent Space's first robotics episode features Dyna Robotics chief scientist Jason Ma on Dyna-2, physical agents designed to work autonomously for an hour. He discusses the company's bet on world action models and physical AI.
MIT researchers look at how machine-learning networks can learn without detailed instructions from a designer. Such networks may help explain how the brain reorganizes itself when it learns.
GitHub describes a context-aware classifier that improves how it detects leaked secrets such as credentials in code. The company reports an 11.5% rise in weighted acceptance and an 86% cut in median scan time.
Google DeepMind has published a post on a robotics version of its new Gemini 4 Argon model, extending the launch beyond chat and coding.