Unverified50% confidenceOpinionExact time
The fine-grained understanding deficiencies of VLMs are believed to be related to the training objective of CLIP-style contrastive learning, as the coarse-grained image-text matching objective cannot prompt models to learn compositional semantics.
1
Sources
50%
Confidence
Long-term
Relevance
9/1/2026
First Seen
Sources
Related Entities
Related Claims
Unverified过拟合是LoRA训练中最常见的失败模式，模型会死记硬背训练图像像素特征而非学习本质特征70% similarUnverified微调VLA模型时通常采用分层学习率策略：视觉编码器使用极小学习率或冻结，语言模型层使用中等学习率，动作专家使用较大学习率68% similarUnverifiedVLM能够根据自然语言描述识别训练数据中从未出现的目标类别，这种能力被称为零样本（zero-shot）或开放词汇（open-vocabulary）识别67% similarUnverifiedThe model training side has tool support from PyTorch, TensorFlow, and tools such as MLflow, Weights & Biases, and DVC, but the degree of standardization on the inference side is notably insufficient.67% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/838969API
curl https://kongchang.com/api/v1/knowledge/claims/838969MCP
get_claim(id=838969)