Unverified50% confidenceFactExact time
扩散语言模型的Transformer通常采用双向注意力，需接收时间步信息（通过正弦编码或AdaLN注入），因此标准预训练语言模型权重不能直接复用
1
Sources
50%
Confidence
Long-term
Relevance
9/1/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedContext forgetting in AI coding tools is a fundamental limitation of Transformer-based large language models, not a flaw specific to any one model71% similarVerified大语言模型本质上是无状态的，不存在跨会话的原生记忆机制，这根植于Transformer架构本身68% similarUnverified主流扩散模型的文字生成依赖CLIP等视觉-语言预训练模型，其训练数据以英文为主，导致中文字符生成困难67% similarUnverified当前几乎所有主流大语言模型都基于Transformer架构67% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/837836API
curl https://kongchang.com/api/v1/knowledge/claims/837836MCP
get_claim(id=837836)