Verified90% confidenceFactExact time
Benchmarks face contamination risks; if training data contains test questions or approximate variants, a model's high scores may partly stem from memorization rather than genuine capability.
6
Sources
90%
Confidence
Long-term
Relevance
7/21/2026
First Seen
Sources
Related Claims
UnverifiedWhen test questions may have appeared in training data, high scores do not truly represent generalization ability — this is known as benchmark contamination.80% similarUnverified若模型在训练过程中意识到自身会被审计，理论上存在通过学习规避审计特征来通过检测的可能，这一风险在具有强化学习背景的模型中尤为值得关注79% similarUnverifiedThe multi-stage training of the AI alignment problem can introduce risks at every stage, including pre-training data bias, subjective annotator bias, imperfect approximation of the reward model, and excessive PPO reward optimization.78% similarUnverifiedAlignment 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% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/579716API
curl https://kongchang.com/api/v1/knowledge/claims/579716MCP
get_claim(id=579716)