Verified75% confidenceCautionExact time
If evaluation benchmarks are highly aligned with training data, surpassing them reflects overfitting-style specialization rather than genuine capability improvement.
3
Sources
75%
Confidence
Long-term
Relevance
9/11/2026
First Seen
Sources
Related Entities
Related Claims
Unverifiedrank值过高、学习率设置不当或训练步数过多是LoRA训练过拟合的常见诱因79% similarUnverifiedIf evaluation samples (including near-duplicate samples) appear in the training data, scores reflect memorization ability rather than genuine skill.79% similarUnverifiedReinforcement learning in the post-training stage can improve answer accuracy on both in-distribution and out-of-distribution tasks, but does not consistently improve the validity of reasoning traces.79% similarVerifiedAlignment training optimizes the model's statistical behavior in known scenarios rather than establishing formally verifiable security boundaries, and may fail when faced with rare instruction combinations or role-playing frameworks outside the training distribution.77% similarUnverified计算最优（Compute-Optimal）并不等于集群最优（Cluster-Optimal），真正决定训练效率的因素远不止浮点运算次数77% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/900001API
curl https://kongchang.com/api/v1/knowledge/claims/900001MCP
get_claim(id=900001)