The study re-examines the claim of lossless speculative decoding in Orthrus, a hybrid autoregressive-diffusion model, under varying numerical precisions. It finds that exact trajectory matching occurs in only 45% of cases with BF16 and 100% with FP32, suggesting that numerical precision significantly affects the model's losslessness.
The chart will appear after repeat observations. The current metric comes from the source.
18 source votesReal observations only. History before source connection is not reconstructed.
This finding highlights the importance of numerical precision in achieving exact output consistency in hybrid models, which is critical for applications requiring deterministic behavior.
The paper introduces Vidu S2, a system with two components: Vidu S2-Avatar for real-time interactive digital characters and Vidu S2-Editing for real-time video editing. It also explores real-time spatial video generation for both components. Compared to its predecessor, Vidu S2-Avatar offers improved real-time 720p video generation, dynamic reference updates, and enhanced instruction following. Vidu S2-Editing enables real-time video stream editing with features like style rendering and background replacement.
5 days agoDaily PapersPhysBrain 1.5 is a unified model for understanding physical environments, generating actions, and predicting future states. It uses a vision-language foundation, encodes language responses and motion as sequences, and is pre-trained on human interaction videos. The model achieves strong performance on embodied understanding benchmarks.
yesterdayDaily PapersZGCM-1 is a 7B parameter foundation model designed for math and agentic search, emphasizing efficiency through architectural and system co-design, progressive curriculum learning, and an AI-native R&D workflow. It demonstrates competitive performance against larger models on mathematical and search tasks, with improvements in training efficiency and open-source availability of training materials.
4 days agoDaily PapersLynnReal-Omni is a native multimodal video generation framework that combines text-to-video, image-conditioned generation, and structural control into a single model. It includes a 32B shared multimodal diffusion transformer and a 27B Flash version for real-time rendering, with a data pipeline and evaluation method designed for video generation tasks.
yesterdayDaily PapersAtria Dawn Preview is a foundation agentic language model designed for scientific research and engineering workflows, trained via a Verifiable Experience Pipeline that connects tool-mediated interactions to executable environments and externally verified outcomes. It achieves competitive performance across 16 benchmarks and shows potential for human-AI collaboration in research and development.
yesterdayDaily PapersRSIAgent is a training-free framework that enables digital agents to autonomously explore and improve themselves in new environments by building and reusing memory. It uses a combination of broad and deep exploration strategies to discover diverse structures and uncover hidden constraints, leading to improved performance on tasks without updating model parameters.
yesterday