Unverified50% confidenceFactExact time
Holding the readout condition number fixed, changing the generative distribution reverses which training coordinate best recovers the known factors.
1
Sources
50%
Confidence
Long-term
Relevance
10/14/2026
First Seen
Sources
Related Entities
Related Claims
Unverified分布漂移是行为克隆的核心缺陷，因训练数据全部来自正确操作轨迹，模型未见过犯错后如何恢复的场景76% similarVerifiedFine-tuning continues to update parameters on top of pre-trained weights through the backpropagation algorithm, adapting the model to a specific domain or task style.74% similarUnverifiedA single-step generation in discrete diffusion can truly match the training data distribution only when the positions it writes to are mutually independent given the already-fixed tokens.74% similarUnverifiedIn the reinforcement learning pipeline, the model continuously generates candidate outputs (rollouts), which are evaluated and scored before being fed back into the training process to update the policy.73% similarUnverified自回归生成系统训练采用Teacher Forcing策略，推理阶段切换为自回归模式后产生分布偏移导致误差累积放大72% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1023667API
curl https://kongchang.com/api/v1/knowledge/claims/1023667MCP
get_claim(id=1023667)