Unverified50% confidenceFactExact time
exposure bias的根源在于训练与推理的分布不匹配，训练时使用teacher forcing策略而推理时模型依赖自身生成的token
1
Sources
50%
Confidence
Long-term
Relevance
8/5/2026
First Seen
Sources
Related Claims
Unverified来自不同厂商、经历不同训练路径的模型持有不同的先验偏见，从而在交叉审查中形成真正的视角差异76% similarUnverified训练数据构成决定模型世界观边界的现象在数字人文领域被称为'语料库偏差'（Corpus Bias）73% similarUnverifiedIf training corpora contain gender, racial, or cultural bias, models will inherit these biases.72% similarUnverifiedSingle judge models have known biases including position bias, length bias, and self-enhancement bias.68% similarVerifiedThere is a gap between benchmark scores and real user experience, and some vendors engage in targeted optimization for specific benchmarks (score gaming) as well as training data contamination that inflates scores.68% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/692490API
curl https://kongchang.com/api/v1/knowledge/claims/692490MCP
get_claim(id=692490)