Unverified50% confidenceFactExact time
OXR's post-training stage employs error-sample-driven iterative data augmentation: samples where the model errs or is uncertain are converted into textual rationales, and the rationales are clustered to drive the generation of targeted training data for the next round.
1
Sources
50%
Confidence
Long-term
Relevance
10/30/2026
First Seen
Sources
Related Entities
Related Claims
Unverified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74% similarUnverifiedMachine learning models may overfit specific training samples and inadvertently reproduce details of the original data when generating content, a phenomenon known as training data memorization.71% similarUnverified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71% similarUnverifiedOverfitting means the model performs perfectly on training data but performs poorly on new data, equivalent to rote memorization rather than learning the underlying patterns.71% similarUnverifiedAutoregressive models suffer from the exposure bias problem, whereby the model sees real prefixes during training but faces self-generated erroneous prefixes during inference, causing errors to accumulate progressively.70% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1115198API
curl https://kongchang.com/api/v1/knowledge/claims/1115198MCP
get_claim(id=1115198)