📑Salesforce: "KV 캐시, 점수 매기지 말고 랜덤으로 버리세요"

이 호에서 다룬 논문

• REASONING KV

  Random Attention: Rethinking KV Cache Eviction for Efficient Reasoning

  Salesforce AI Research, UIUC · KV Cache Eviction, Efficient Reasoning, Random Selection

• RL

  StudentSim: Training LLM-based Student Simulators

  Microsoft Research, UIUC · Student Simulator, Personalized Tutoring, Reward Model

• VISION

  Qwen-Drive-1.0: An Initial Step towards a Vision-Language Foundation Model for Autonomous Driving

  Alibaba (Qwen Team) · Autonomous Driving, Vision-Language Model, Motion Planning

• AI RESEARCH

  Compile by Training: Turning Natural-Language Specifications into Local Neural Functions

  University of Waterloo, Harvard · Neural Compilation, Small Adapter, Local Inference

• AGENTS

  Terminal-Universe: Turning Agent Trajectories into Scalable Terminal Environments

  Alibaba (Qwen Team) · Agent Environment, Trajectory Reconstruction, Post-Training

• GENERATIVE

  LLaDA-Image: Building Strong Image Generators with Fully Open Training Recipes

  Ant Group (inclusionAI) · Diffusion Transformer, Open Recipe, Unified Generation

• VISION

  Does On-Policy Distillation Really Distill? From Noisy Teacher to Self-Improvement

  Purdue University · On-Policy Distillation, Supervision-Free RL, Entropy-Adaptive

• ROBOTICS

  RoboTok: An Internet-Scale Data Engine for Human Demonstration Retrieval and Dexterous Manipulation Learning

  Rice University · Robot Learning, Web Video Retrieval, Dexterous Manipulation

• NLP

  LatentPress: Context Compression Beyond Text and Vision

  LinkedIn · Context Compression, Soft Tokens, Long Context

• VISION

  DreamX-Creator: Democratizing Native Audio-Video Generation at 2K Resolution

  Alibaba (AMAP) · Audio-Video Generation, Joint Denoising, 2K Refinement