Unverified50% confidenceSolutionExact time
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
1
Sources
50%
Confidence
Long-term
Relevance
8/6/2026
First Seen
Sources
Related Claims
UnverifiedVLA models unify visual perception, language understanding, and action generation within a single model framework, and their training is highly dependent on large-scale, high-quality human manipulation demonstration data.74% similarUnverifiedLLaVA的训练分为两个阶段：第一阶段是预训练对齐（冻结视觉编码器和语言模型，仅训练投影层），第二阶段是指令微调（解冻语言模型进行端到端训练）73% similarUnverifiedThe Cosine Annealing learning rate scheduling strategy is particularly suitable for LoRA training, where a high learning rate in the early stages of training rapidly captures features and then shrinks in the later stages to prevent overfitting.70% similarUnverified通过LoRA等轻量化微调技术，可以用少量特定数据对模型进行定向训练，生成特定类型内容70% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/696669API
curl https://kongchang.com/api/v1/knowledge/claims/696669MCP
get_claim(id=696669)