Unverified50% confidenceSolutionExact time
Numerical consistency testing can be systematized into a regression test suite that computes the maximum absolute error and KL divergence of per-token log probabilities from both sides after each engine version upgrade, with a tolerance threshold (e.g., KL < 1e-4) that blocks training when exceeded.
1
Sources
50%
Confidence
Long-term
Relevance
9/12/2026
First Seen
Sources
Related Entities
Related Claims
Unverifiedpass@k指标衡量模型生成k个候选代码片段时至少有一个能通过全部测试用例的概率74% similarUnverifiedIf training error can precisely track test error, it becomes possible to achieve an optimal stopping strategy without a validation set and more efficient hyperparameter tuning.72% similarUnverified当比较两个具体训练模型的预测行为差异时应固定种子使用同一测试集运行McNemar检验，当评估方法整体优越性时应用多种子重复实验配合配对t检验或非参数检验71% similarUnverified通过回归测试主要验证功能正确性和输出一致性，不覆盖并发竞态条件、极端负载行为或长时间运行稳定性70% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/916509API
curl https://kongchang.com/api/v1/knowledge/claims/916509MCP
get_claim(id=916509)