Celeris is developing faster large language models by designing for speed from inception, focusing on diffusion-based approaches that enable parallel token generation. The company combines sequential and parallel decoding within hybrid architectures, leveraging research showing diffusion models can achieve 27.6× throughput improvements and better reasoning capabilities than autoregressive models.
Cortee is a diffusion model-based tool that helps users write documentation quickly by converting brief notes into full paragraphs, with edits applied live to documentation pages. The service prioritizes user control and data privacy, storing documents in the EU while offering data export and deletion options.
This paper challenges common assumptions about why generative models succeed in robotic control, finding that their advantage comes from iterative computation with proper supervision and stochasticity rather than multi-modal distribution capture. A lightweight two-step regression-based policy matches or exceeds performance of more complex generative control policies, suggesting the distribution-fitting component is less critical than believed.
Google Research's Retrieve-for-Train framework uses offline reinforcement learning to train lightweight diffusion models for efficient AI search result generation, eliminating expensive inference-time reasoning. Instead of forcing language models to perform costly query decomposition during search, the approach pre-trains models to return coherent, diverse, and complementary result sets that satisfy set-level properties like coverage and complementarity.
Researchers released an open-source frontier-level world model trained on the Dreamer4 paper using the CoinRun game, featuring a block-causal transformer architecture with a tokenizer and dynamics model. The project demonstrates scaling laws where compute-optimal parameters and data points scale with the square root of compute, and combines world prediction with policy learning through diffusion-based latent state generation.