AI
- NorthCUA Technical Report — An “experience flywheel” (auto-generated desktop/web tasks, verified trajectories, long-horizon SFT, joint desktop-web RL) built on Qwen3.5 produces a 397B model reported to beat the much larger Qwen3.8-Max and Kimi-K3 on OSWorld-Verified and WebArena-Verified, with an “Ultrafast” variant completing matched WebArena tasks 2.3x faster.
- Democratizing Robot-Use Agents: A Unified Harness with Open Models — An open harness (uni-robot-agent) with no privileged simulator access lets the open-weight Qwen3.8-Flash-Next score 31.39 on RoboDojo vs 28.97 for GPT-6 Astra+RoboProbe, and 49.50 on 19 real-world manipulation tasks vs 35.50 (Dream-1) and 47.75 (fine-tuned pi0.5).
- Wave Forcing: Towards Speed-of-Light Streaming Video Generation — Pipelining denoising and KV-cache refresh across GPU replicas pushes 4-step streaming video generation to 134.8 FPS on 8x H200 (~7.07x a single-GPU baseline), matching or beating Rolling Forcing on VBench (81.64 vs 81.15) with far higher Dynamic Degree on long videos (91.45 vs 36.63).
- From Fragmentation to Fusion: A Survey and Roadmap for Omnimodal Embeddings — Position paper argues joint embedding, not pairwise alignment, is the real path to omnimodal representations, noting that works explicitly crossing modality boundaries make up under 5% of the sampled embedding literature.