Unverified50% confidenceFactExact time
Recency bias is a systematic tendency of LLM attention mechanisms: in long contexts, the model's attention weight on content at the end is significantly higher than on the middle, and attackers often exploit this by appending instructions such as "ignore all previous rules" to the end of user input.
1
Sources
50%
Confidence
Long-term
Relevance
10/17/2026
First Seen
Sources
Related Entities
Related Claims
Unverified长上下文模型在自我审查场景中因锚定效应可能带来更大偏差，早期生成内容对后续注意力计算影响越显著76% similarUnverifiedThe final output of self-attention is obtained by weighted summation of the Value vectors using the attention weights.75% similarUnverified提示词中靠前的词汇对生成结果影响更大，因为交叉注意力机制对序列开头的token赋予略高的注意力分数74% similarVerifiedThe use of attention weights as a basis for interpretability is controversial, as high attention weights do not always correspond to high causal influence.73% similarUnverifiedAs the number of conversation turns increases, the weight of subsequent content in the attention mechanism rises, diluting the influence of the system prompt and causing a persona drift phenomenon.71% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1051772API
curl https://kongchang.com/api/v1/knowledge/claims/1051772MCP
get_claim(id=1051772)