📑Meta: "강화학습이 AI를 편식하게 만든다"

이 호에서 다룬 논문

• VISION

  LoopVL: Recurrent Visual Intelligence

  Baidu, Renmin University · Loop Transformer, Vision-Language, Weight Sharing

• AGENTS

  Sharpening Tax in Post-Training

  Meta (Superintelligence Labs) · RL Post-Training, pass@K, Agentic Task

• AGENTS

  Beyond Memory: Harnessing Long-Horizon Agents with Explicit Belief States

  Alibaba Group, Nankai University · LLM Agent, Belief State, Long-Horizon

• ROBOTICS

  Native Action-Prior Learning from Videos for World Action Models

  Meta AI · World Action Model, Observation-only Video, Robot Control

• AGENTS

  VeriHarness: Scaling Agentic Verification for Long-Horizon Tasks

  Google Cloud AI Research, University of Cambridge · Agentic Verification, Long-Horizon, Self-Improvement

• WORLD MODEL

  Spatial Memory Intelligence: Endowing World Models with Understanding-Driven Long-Term Memory

  Alibaba Group, CUHK-Shenzhen · World Model, Long-Term Memory, MLLM

• VISION

  ProAR: Learning Prospective Reasoning with Autoregressive Video Models

  Microsoft, Hong Kong Polytechnic University · Autoregressive Video, Prospective Reasoning, Goal-Frame

• ROBOTICS

  EyeRobot 2.0: Active Gaze for Precise Manipulation without Wrist Cameras

  Amazon (FAR), UC Berkeley · Active Gaze, Bimanual Manipulation, Foveal Vision

• AGENTS AI

  Self-Supervised Scaling of Terminal Environments for Scientific Domains

  Tencent (HY LLM Frontier) · Terminal Agent, Self-Supervised, Scientific Domain

• VISION

  Breaking the Uniformity Trap: Scaling Video Diffusion Model via SplitMoE

  ByteDance, UCAS · Mixture-of-Experts, Video Diffusion, Routing