Unverified50% confidenceBenchmarkExact time
By the middle layers of the model, the norm of the register vectors grew to 4–13 times that of the image tokens.
1
Sources
50%
Confidence
Long-term
Relevance
9/11/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedThe ratio of input to output tokens surged from 4.52:1 to 13:1.71% similarUnverified靶向数据工程中，针对性数据占总微调数据的 5%-15% 通常能取得良好效果，过高比例可能导致模型在其他维度退化63% similarUnverifiedEarly models had a context length of approximately 4,096 tokens, whereas modern mainstream models have expanded to hundreds of thousands or even millions of tokens.63% similarUnverified推荐为每个人注册5-10张不同条件下的照片，使用多模板平均向量可将识别准确率提升2-5个百分点62% similarUnverified同一段文字在不同模型中消耗的 token 数量可能相差 10%-30%62% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/902317API
curl https://kongchang.com/api/v1/knowledge/claims/902317MCP
get_claim(id=902317)