📑Huawei: "엔비디아 없이 DeepSeek-V4를 길렀습니다"

이 호에서 다룬 논문

• AGENTS

  AREX: Towards a Recursively Self-Improving Agent for Deep Research

  Independent · Recursively Self-Improving, Deep Research Agent, Long-horizon RL

• TRAINING

  SLAI T-Rex: Full-Parameter Post-training of the DeepSeek-V4 Family on Ascend SuperPOD

  Huawei · Ascend NPU, Trillion-parameter MoE, Full-parameter Post-training

• VISION

  SANA-Video 2.0: Hybrid Linear Attention with Attention Residuals for Efficient Video Generation

  NVIDIA · Hybrid Linear Attention, Video Diffusion, Efficient Generation

• VISION

  Robostral Navigate

  Mistral · Monocular Navigation, Vision-Language Model, Cross-embodiment

• AGENTS

  Tencent WorkBuddy Bench: A Multi-Domain Coding-Agent Benchmark with Contamination-Resistant Task Construction

  Tencent · Coding Agent Benchmark, Contamination Resistance, Multi-domain Evaluation

• AGENTS

  LLMs Get Lost in Evolving User Intent

  Microsoft · Evolving Intent, Multi-turn Interaction, Collaborative Agents

• WORLD MODEL AI

  Streaming Multi-Agent Autoregressive Diffusion Model with World State Registers

  Adobe, UCLA · Multi-agent World Model, World State Registers, Streaming Diffusion

• VISION

  Visual Contrastive Self-Distillation

  University of Maryland · Self-Distillation, Vision-Language, On-policy Learning

• VISION

  Show, Don't Tell: Evaluating Spatial Cognition in Generative Pixels Rather Than LLM Text

  Independent · Spatial Cognition, Image-generation Evaluation, Pixel-space Answers

• VISION

  Self Gradient Forcing: Native Long Video Extrapolation

  Independent · Autoregressive Video Diffusion, Self Forcing, Long-video Extrapolation