Unverified50% confidenceFactExact time
Recurrent-based models are prone to hidden state collapse and gradient instability when trained via backpropagation through time (BPTT).
1
Sources
50%
Confidence
Long-term
Relevance
10/9/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedEarly LLMs were prone to stalling on long-horizon tasks, falling into repetition, losing direction, or being unable to recover from errors.73% similarUnverified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72% similarUnverifiedBehavior Cloning (BC) suffers from the covariate shift problem: during training the model only sees the state distribution of expert trajectories, and at deployment small errors can drive the robot into states not covered by the training set, leading to cascading failures.71% 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70% similarVerifiedGANs have inherent drawbacks such as training instability and susceptibility to mode collapse, which prompted the research community to shift toward diffusion models that generate images through a gradual denoising Markov chain process.70% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/990497API
curl https://kongchang.com/api/v1/knowledge/claims/990497MCP
get_claim(id=990497)