📑Google: "스스로 진화한 에이전트, 토큰 30% 덜 쓴다"

이 호에서 다룬 논문

• VISION

  WorldCrafter: Consistent Video World Model with Implicit 3D-aware Memory

  Tencent (ARC Lab) · Video World Model, 3D-aware Memory, Camera Control

• AGENTS

  RRSI: Regularized Recursive Self-Improvement of Agent Harnesses

  Google Research · Agent Harness, Recursive Self-Improvement, Regularization

• AI RESEARCH

  GameHorizon Suite: Multi-Horizon Data and Evaluation in Gameplay

  Tencent · Gameplay Benchmark, Multi-Horizon, Dataset

• VISION

  OmniVChat: Synthesizing, Benchmarking, and Training for Native Audio-Visual Dialogue

  Alibaba (Qwen) · Omni Model, Audio-Visual Dialogue, RL

• VISION

  Paint-Anything: Unified Any-Color Control for Image Generation and Editing

  ByteDance (Seed) · Image Generation/Editing, Color Control, Hex Prompt

• SAFETY

  onPanda: Efficient Annotation of On-Policy Alignment Data for LLMs and Agents via Token-Level Correction

  StepFun · Alignment Data, Token-Level Correction, On-Policy Data

• WORLD MODEL

  Training Object Permanence in World Models

  Johns Hopkins, Oxford, Stanford 외 (다기관) · World Models, Object Permanence, Physical Reasoning

• ROBOTICS

  Grounded Action Model: 3D Grounding as a Foundation for Robotics

  University of Washington, Allen Institute for AI (AI2) · Robotics, 3D Grounding, VLA/World-Action Models

• RL

  RULER: Instance-aware Rubric Rewards for SVG Generation

  National University of Singapore (Show Lab) · SVG Generation, Rubric Rewards, RL

• NLP

  Your Transformer Can Hold Two Thoughts at Once: Evidence of Linear Superposition in LLMs

  AIRI · Interpretability, Linear Superposition, LLM