Unverified50% confidenceBenchmarkExact time
For a published 406M-parameter Fusion-in-Decoder specialized model, the ROUGE-1 score dropped sharply by 6.30 points when the input was switched from truncated long text to 2,000-word retrieved segments.
1
Sources
50%
Confidence
Long-term
Relevance
9/23/2026
First Seen
Sources
Related Entities
Related Claims
Unverified早期Stable Diffusion系列以CLIP为文本编码器，对超过77个Token的提示词处理能力受限68% similarUnverified在Stable Diffusion中，CLIP文本编码器将自然语言切分为最多77个token，每个token映射为768维或1024维嵌入向量66% similarUnverifiedThe classic Soundex algorithm produces only 4-character codes, with an encoding space of at most approximately 8,000 distinct values, leading to high collision rates when dealing with large-scale vocabularies.66% similarUnverifiedCLIP文本编码器的标准上下文窗口长度约为77个Token，超出部分的语义信息会被截断或降权处理65% similarUnverifiedIn large model invocations, text and code in both input and output are converted into token counts.65% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/942872API
curl https://kongchang.com/api/v1/knowledge/claims/942872MCP
get_claim(id=942872)