📑Shanghai AI Lab: "학습 데이터 37%만 썼는데 성능이 같다니까?"

이 호에서 다룬 논문

• AI

  > Intern-S2-Mobius: Foundation Model with Decoupled Knowledge and Reasoning

  Shanghai AI Lab, CUHK · Knowledge-Reasoning Separation, Mixture-of-Experts, Inference Speedup

• AI RESEARCH

  > VoiceChat-TTS: A Low-Latency Continuous Speech Synthesis Model for Interactive Agents

  NVIDIA · Continuous TTS, Barge-in, Low-Latency

• AGENTS

  > ScienceFlow: A long-horizon agent for ML research, scientific discovery and beyond

  Huawei Noah's Ark Lab · Autonomous Research Agent, Long-Horizon, State Recovery

• AI RESEARCH

  > ASSERT: A Measurement Pipeline for GenAI Audits

  Microsoft Research · GenAI Audit, Measurement Validity, Specification-Driven

• VISION

  > Can We Defend Against AI-Generated Video Attacks on Real-World Crisis Events? A Systematic Evaluation of Detectors, Generators and Social Dissemination

  Zhejiang University, NUS · Deepfake Video Detection, Crisis Misinformation, Social Dissemination

• VISION

  > Self-Supervised Visual On-Policy Distillation

  Oxford, UC San Diego · Self-Supervised Distillation, Visual Augmentation, On-Policy Learning

• AGENTS

  > AutoDesign: Meta-Harness Optimization for Long-Horizon Agentic Design

  Independent · Agentic Design, Paper-to-Poster, Meta-Optimization

• WORLD MODEL

  > Marionette: Predicting World States, Rendering Geometry, Painting Appearance

  Independent · World Model, Articulated Character, Explicit State

• AI RESEARCH

  > Improving the matrix multiplication exponent with modern optimization and AlphaEvolve

  Google DeepMind, MIT · Matrix Multiplication Exponent, AlphaEvolve, Combinatorial Optimization