GLM-5V-Turbo

8 atoms · 跨 1 天 · 首见 2026-05-07 · 最近 2026-05-07

三色: 🟦 fact 8 · 🟥 take 0 · stance ▲0/▼0/◆0

来源: X 8

时态: fresh:8

标签: 好数字:5 · 好观点:3

事实 (fact) (8)

• 2026-05-07 · CogViT Vision Encoder built with dual-teacher distillation
  CogViT Vision Encoder Built with dual-teacher distillation: SigLIP2 for semantics, DINOv3 for texture.
  tags: 好数字 · → daily
• 2026-05-07 · Multimodal Multi-Token Prediction (MMTP) uses shared <image> token
  MMTP uses a shared <image> token, removing the need to propagate visual embeddings across pipeline stages and improving training stability.
  tags: 好数字 · → daily
• 2026-05-07 · joint RL across 30+ task categories yields consistent gains
  Joint RL across 30+ task categories yields consistent gains with weaker cross-domain interference than SFT.
  tags: 好数字 · → daily
  value: qty=30+ task categories
• 2026-05-07 · multimodal RL infrastructure rebuilt on four axes
  Infrastructure rebuilt along four axes: unified task and reward abstraction, full-pipeline asynchrony, fine-grained memory management for vision modules, and topology-aware partitioning for variable-length visual inputs.
  tags: 好数字 · → daily
• 2026-05-07 · Multimodal RL infrastructure built with QK-Norm for attention stability at scale
  QK-Norm for attention stability at scale.
  tags: 好数字 · → daily
• 2026-05-07 · perception remains foundational for agentic tasks
  Perception remains foundational. Many high-level failures begin with the model not seeing accurately enough.
  tags: 好观点 · → daily
• 2026-05-07 · hierarchical optimization works better than monolithic end-to-end training
  Hierarchical optimization works better than monolithic end-to-end training. Distributed optimization across perception, single-step actions, and trajectory planning yields more stable agentic capability.
  tags: 好观点 · → daily
• 2026-05-07 · end-to-end agent tasks need clear specification, reliable verification, and controlled evaluation
  End-to-end agent tasks need clear specification, reliable verification, and controlled evaluation. Without these, realistic settings are too open-ended to produce reusable optimization signals.
  tags: 好观点 · → daily

时间轴 (近 20)

• 2026-05-07 · fact · perception remains foundational for agentic tasks · → daily
• 2026-05-07 · fact · hierarchical optimization works better than monolithic end-t · → daily
• 2026-05-07 · fact · end-to-end agent tasks need clear specification, reliable ve · → daily
• 2026-05-07 · fact · CogViT Vision Encoder built with dual-teacher distillation · → daily
• 2026-05-07 · fact · Multimodal Multi-Token Prediction (MMTP) uses shared <image> · → daily
• 2026-05-07 · fact · joint RL across 30+ task categories yields consistent gains · → daily
• 2026-05-07 · fact · multimodal RL infrastructure rebuilt on four axes · → daily
• 2026-05-07 · fact · Multimodal RL infrastructure built with QK-Norm for attentio · → daily

← 实体目录 · 系统日志