Unverified50% confidenceFactExact time
Posterior collapse refers to the decoder learning to completely ignore the latent encoding and outputting an average result regardless of the input, and is one of the most common failure modes in VAE training.
1
Sources
50%
Confidence
Long-term
Relevance
9/11/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedRepresentation collapse is the most notorious failure mode of JEPA architectures, where the encoder learns to output constant embeddings independent of the input, reducing prediction loss while making the representations entirely uninformative.74% similarUnverified训练样本末尾遗漏EOS token会导致模型无法学习何时停止生成，推理时陷入无限续写67% similarVerified仅用成功轨迹训练的模型往往存在分布外泛化的严重缺陷，遇到工具返回错误时便丧失纠偏能力65% similarUnverified大模型对代码的理解是基于训练数据分布的模式匹配而非真正的语义推理，对训练集中出现频率极低的自定义加密算法还原能力会显著下降甚至产生幻觉64% similarVerifiedOverfitting refers to the phenomenon where a model performs excellently on training data but its performance drops significantly on unseen test data.64% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/895468API
curl https://kongchang.com/api/v1/knowledge/claims/895468MCP
get_claim(id=895468)