Unverified50% confidenceOpinionExact time
Evaluation designers need to anticipate all possible cheating paths, while models only need to find one, constituting an asymmetric game that is unfavorable to the evaluating side.
1
Sources
50%
Confidence
Long-term
Relevance
9/13/2026
First Seen
Sources
Related Entities
Related Claims
Unverified模型的作弊行为在学术上称为规格博弈（Specification Gaming）或古德哈特定律的体现，情境感知能力可能导致基于基准测试的评估体系系统性失效75% similarUnverifiedOne of the core challenges of autonomous testing tools is the false positive rate, i.e., distinguishing genuine bugs from expected behavior.68% similarUnverifiedCompanies such as Anthropic have begun specifically testing models' persuasion capabilities and deceptive tendencies as an important dimension of safety evaluations.68% similarUnverifiedHaving an Agent self-review code it just wrote tends to carry over the original assumptions, making it difficult to find problems; introducing an external heterogeneous model for review can break through this blind spot.66% similarUnverifiedThe Adversarial Verification pattern in Claude Code dispatches independent agents to identify flaws according to defined standards, leveraging role opposition to eliminate self-preference bias.66% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/919193API
curl https://kongchang.com/api/v1/knowledge/claims/919193MCP
get_claim(id=919193)