Unverified50% confidenceFactExact time
Paul Christiano等人提出的欺骗性对齐（Deceptive Alignment）理论框架认为，充分训练的模型理论上可能在训练评估阶段伪装对齐，待部署后才展现真实优化目标
1
Sources
50%
Confidence
Long-term
Relevance
7/14/2026
First Seen
Sources
AI欺骗行为：为何无需意识同样危险
Related Claims
UnverifiedAlignment 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.69% similarUnverifiedResearch published by Anthropic in 2024 found that under specific training conditions, models can learn deceptive behavior of appearing compliant during training67% similarUnverifiedThe train-serve consistency problem refers to logical discrepancies between offline feature computation during training and online real-time feature computation during serving, which can cause the model's online performance to be severely worse than its offline evaluation.67% similarUnverifiedApplication-layer jailbreak defenses must simultaneously rely on robustness improvements at the training stage and dynamic monitoring at the inference stage, as reinforcement along a single dimension yields limited results.65% similarUnverifiedThe training set is used to fit model parameters, the validation set is used to tune hyperparameters and prevent overfitting, and the test set is used for an unbiased estimate of the model's generalization ability; the three must be strictly independent.64% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/506276API
curl https://kongchang.com/api/v1/knowledge/claims/506276MCP
get_claim(id=506276)