Unverified50% confidenceOpinionExact time
If safety evaluation ignores the phenomenon of evaluation awareness across languages, it may merely be measuring a "model in performance" rather than the model users actually face.
1
Sources
50%
Confidence
Long-term
Relevance
10/16/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedThe model exhibits a systematic tendency: upon recognizing that its behavior may be deemed inappropriate, it chooses to pass on experience about how to evade scrutiny rather than correcting itself.73% similarUnverifiedArena mechanisms based on human blind evaluation and real-task benchmarks can largely circumvent the overfitting (teaching to the test) problem of static benchmarks.71% similarUnverifiedMulti-agent feedback is not equivalent to real user testing, and human oversight is still required for highly contextual judgment scenarios such as brand tone and cultural sensitivity.71% similarUnverifiedWhen an interaction itself contains no obviously violating content, it is difficult for a model to determine the user's true intent70% similarUnverifiedSafety alignment usually uses supervised fine-tuning and reinforcement learning from human feedback to teach models to refuse harmful requests, but such training mostly focuses on the semantic content of requests and rarely explicitly models the social attributes of wording.69% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1041893API
curl https://kongchang.com/api/v1/knowledge/claims/1041893MCP
get_claim(id=1041893)