Unverified50% confidenceOpinionExact time
The attention mechanism of LLMs is activated by tokens present in the input; tokens that do not exist produce no activation signal, so this blind spot does not disappear regardless of model scale.
1
Sources
50%
Confidence
Long-term
Relevance
9/16/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedLLM无法从机制层面区分哪些token是指令、哪些是数据，因为所有输入在进入注意力层前都被统一转换为嵌入表示79% similarUnverifiedLLM在架构层面采用Transformer的注意力机制，每次推理在当前上下文窗口内独立完成，不存在跨会话的持久化状态76% similarUnverifiedBecause the attention-mechanism-based architecture of LLMs struggles to precisely distinguish between data to be processed and instructions to be followed, there is currently no perfect purely model-level solution to prompt injection attacks.74% similarVerifiedThe underlying mechanism of LLM hallucination is that the model is optimized to generate text that 'sounds plausible' rather than 'factually accurate,' and the model has no independent knowledge retrieval module.72% similarPartially VerifiedEach LLM inference is stateless and constrained by the limited context window; the model has no persistent memory of its own.70% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/926646API
curl https://kongchang.com/api/v1/knowledge/claims/926646MCP
get_claim(id=926646)