Unverified80% confidenceOpinionTime unknown
自训练循环存在模型坍缩、错误放大和安全评估时间窗口压缩等潜在风险
1
Sources
80%
Confidence
Long-term
Relevance
6/3/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedThe true difficulty of model fine-tuning lies in the construction and cleaning of high-quality training data, preventing catastrophic forgetting, and inference deployment after quantization.73% similarUnverified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73% similarVerifiedAfter geometric transformation, the transformation parameters must be recorded and the predicted boxes must be inverse-transformed to align them; ignoring this step will cause semantic misalignment in the consistency loss and harm training effectiveness73% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/39215API
curl https://kongchang.com/api/v1/knowledge/claims/39215MCP
get_claim(id=39215)