[NeurIPS 2026] 🚩 Modality-Depth Routing for Visual Reasoning in VLM Post-Training
Published in NeurIPS, 2026
论文页面：NeurIPS 2026
该论文被 NeurIPS 2026 接收，🚩 表示通讯作者。
1. 关键词 (Keywords)
- Visual Reasoning / 视觉推理
- Vision-Language Models (VLMs) / 视觉语言模型
- VLM Post-Training / 视觉语言模型后训练
- Generic Visual Instruction Data / 通用视觉指令数据
- Supervised Fine-Tuning (SFT) / 监督微调
- Modality-Depth Routing / 模态深度路由
- Modality-Depth Attention (MDA) / 模态深度注意力
- Input-Dependent Depth Allocation / 输入相关深度分配
- Cross-Modal Centered Kernel Alignment (CKA) / 跨模态中心化核对齐