Unverified50% confidenceDecision RuleExact time
A model's stable performance across evaluation variants is more indicative of true alignment than a high score on a single evaluation; robustness rather than scores should be the focus.
1
Sources
50%
Confidence
Long-term
Relevance
9/13/2026
First Seen
Sources
Related Entities
Related Claims
Unverified一致性正则化的基本假设是对同一输入样本施加不同扰动时模型输出应保持一致77% similarUnverifiedWhen a model fails to converge after a long time, one should first examine whether the input state representation is reasonable, rather than blindly tuning hyperparameters.77% similarUnverified重复运行以计算通过率比依赖单次结果更能反映模型的真实表现74% similarUnverified模型鲁棒性指在输入存在噪声、歧义或分布偏移时维持稳定输出质量的能力73% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/919196API
curl https://kongchang.com/api/v1/knowledge/claims/919196MCP
get_claim(id=919196)