Unverified50% confidenceFactExact time
Language models assign higher weight to content in the context that is closer to the end (recency attention/recency bias).
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78% 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75% similarUnverifiedRecency 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.74% similarUnverifiedThe attention mechanism of large language models may be drawn by the user's specific question, the names of political entities in the context, and political commentary in the training data, thereby probabilistically overpowering the influence of the system prompt.73% similarUnverifiedDifferent attention heads in multi-head attention tend to capture different types of semantic relationships, with some focusing on syntactic dependencies and others on coreference relations.72% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1054476API
curl https://kongchang.com/api/v1/knowledge/claims/1054476MCP
get_claim(id=1054476)