Unverified50% confidenceFactExact time
少样本提示词本身长得多，单纯的长度差异就足以移动模型的表征，过去测到的漂移混杂了与示例内容无关的长度噪声
1
Sources
50%
Confidence
Long-term
Relevance
9/16/2026
First Seen
Sources
Related Entities
Related Claims
Unverified如果设定的时长超过了实际提供的内容量，模型会用无意义的杂音去填补空白76% similarUnverified高偏差意味着模型过于简单导致欠拟合，高方差意味着模型学入噪声导致过拟合，两者往往此消彼长76% similarUnverified扩散模型的迭代去噪过程对数值误差敏感，每步的微小误差会累积并影响最终图像，采样步数越多量化误差影响越大74% similarUnverified主体漂移本质上是模型在时序一致性建模上的不足，源于扩散模型每帧去噪引入随机性叠加时序注意力对长程依赖捕获能力有限74% similarUnverified上下文越长，模型推理速度越慢，即存在上下文长度与推理性能之间的权衡74% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/924778API
curl https://kongchang.com/api/v1/knowledge/claims/924778MCP
get_claim(id=924778)