Graph Learning · LLM × Graph · Multi-Agent · Science
Showing 0 papers for 2026-02-21
Unified Latents (UL) learns latent representations that are jointly regularized by a diffusion prior and decoded by a diffusion model. By linking the encoder’s output noise to the prior’s minimum noise level, UL yields a simple training objective that provides a tight upper bound on latent bitrate. On ImageNet-512, UL achieves a competitive FID of 1.4 with high reconstruction quality (PSNR) while using fewer training FLOPs than Stable Diffusion latent models; on Kinetics-600 it sets a new state-of-the-art FVD of 1.3.
GUI-Owl-1.5 is the latest native GUI agent model with instruct and thinking variants across multiple sizes (2B/4B/8B/32B/235B) and platform support (desktop, mobile, browser, and more) to enable cloud–edge collaboration and real-time interaction. It achieves state-of-the-art results on more than 20 GUI benchmarks, including 56.5 on OSWorld, 71.6 on AndroidWorld, and 48.4 on WebArena for GUI automation; 80.3 on ScreenSpotPro for grounding; and 47.6 on OSWorld-MCP for tool‑calling tasks.
We study three key questions about trainable sparse attention in diffusion-model acceleration: (1) when do the common masking rules Top-k and Top-p fail, and how can we avoid these failures; (2) why can trainable sparse attention achieve higher sparsity than training-free methods; (3) what are the limitations of fine-tuning sparse attention with the diffusion loss, and how can we address them. Based on this analysis, SpargeAttention2 proposes methods to improve sparsity while preserving generation quality.
Frontier AI Risk Management Framework in Practice presents a comprehensive risk analysis of frontier AI, focusing on rapidly evolving LLMs and agentic AI. It provides a granular assessment across five dimensions: cyber offense, persuasion and manipulation, strategic deception, uncontrolled AI R&D, and self-replication, with further refinements to components and guidance for practitioners.
Arcee Trinity Large Technical Report describes Arcee Trinity Large, a sparse Mixture-of-Experts model with 400B total parameters and 13B tokens activated per step; also reports on Trinity Nano (6B total, 1B activated) and Trinity Mini (26B total, 3B activated). The architecture features interleaved local and global attention, gated attention, depth-scaled sandwich norm, and sigmoid routing for MoE. For Trinity Large, a new MoE load-balancing strategy is introduced.