A training-free decoder exploits spatial and temporal stability in diffusion language models to deliver up to 14x throughput gains without hurting accuracy.
A browser-based voice synthesizer skips samples and neural nets entirely, generating singing, speech, and beatboxing from a simulated airway written in Rust.
Researchers carve out an 'eager' fragment of timeline-based planning that skips the exponential determinization step and directly synthesizes strategies.
A Google DeepMind team proves that two natural ways to train language models on their own decoded outputs are broken by construction, and concludes that teacher forcing remains the right tool for supervised training.
NVIDIA's 117M-parameter Déjà View loops one transformer block to beat billion-parameter 3D reconstruction baselines while letting you dial compute at inference.
Google Research unveils SensorWM, a world model trained on 2.28M person-weeks of Fitbit data that simulates 12-week what-if habit trajectories for personalized health coaching.
Google opens Managed Agents on the Gemini API to bring-your-own sandbox runtimes, with seven launch partners including Vercel, Modal, Cloudflare and E2B.
Zed v1.28 lands with optional background Git auto-fetching, a resizable Refs column in the Git Graph, and mouse-wheel font zoom in diffs and terminals.
Perplexity's follow-up red-team study shows that Capture the Flag framing misdirects agents away from the actual security boundary, and cyber-tuned models don't always win.
Google's new Gemini Decision 1 model scores up to 1,000 multimodal questions in a single forward pass, returning calibrated probabilities instead of text.