📑NVIDIA: "GPU 1장이면 753B도 서빙됩니다"

이 호에서 다룬 논문

• RL

  EnvHarness: Awakening Static Worlds for Agent Learning

  Google · Adaptive Environment, Agent Training, Co-Evolution

• AI RESEARCH

  FreeToken: Efficient Edge-Native MoE Serving with Bandwidth-Adaptive Execution

  NVIDIA, MIT, UC Berkeley · Edge Inference, MoE Serving, Bandwidth-Adaptive

• RL

  SPADE: Self-Play in Adaptive Synthetic Executable Environments

  Meta, University of Washington, MIT · Self-Play RL, Environment Design, Open-Ended Learning

• AGENTS

  LEGO-RL: Harness-Native Reinforcement Learning for Coding Agents

  Huawei, CUHK · Coding Agent RL, Harness Alignment, Reward Hacking Defense

• NLP API

  Ventor-QTest: Threat-Model-Driven Verification of Vendor-Hosted LLM APIs

  Tencent · LLM API Audit, Black-Box Verification, Fidelity Loss

• VISION

  SemComp-Bench: Benchmarking Semantic Task Completion in Video Generation

  USTC · Video Generation, Task Completion, Semantic Grounding

• RL

  Learn What's Left, Not What's Mastered: Saturation Aware Advantage Reweighting for Multi-Reward Policy Optimization

  UCSD, Northeastern University · Multi-Reward RL, Saturation-Aware, Adaptive Optimization

• AGENTS

  FACET: Preserving Source Intent and Executable State in Terminal Task Synthesis

  USTC · Terminal Agent Training, Task Synthesis, Executable Verification

• GENERATIVE PLC

  SemaPLC: A Project-Grounded, Verification-Gated Agent Harness for PLC Code Generation

  Midea AI · PLC Code Generation, Verification-Gated, Industrial Automation

• AGENTS

  Co-RL: Unsupervised Reasoning Emerges from Diverse Cohort in Multi-agent RL

  UCSD · Unsupervised RL, Multi-Agent, Diverse Cohort