Unverified50% confidenceFactExact time
The researchers constructed two variants of the same base teacher model: one tending to accept harmful requests (harm-accepting) and the other tending to avoid harmful content (harm-avoiding), with the two differing only in safety tendency rather than task ability.
1
Sources
50%
Confidence
Long-term
Relevance
10/14/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedWhen the goal is to expose true differences between models, they should be forced to respond quickly rather than deliberate, and the possibility of mutual imitation should be eliminated through mechanism design.64% similarUnverifiedThe vulnerability-modeling mechanism is consistent with Albert Bandura's social learning theory, which holds that the transmission of behavioral norms depends on observable role models rather than preaching.64% similarUnverifiedWhen the teacher model is a competitor's closed-source frontier model, distillation raises issues of intellectual property and terms of service.64% similarUnverifiedFull-duplex models are highly sensitive to how instructions are presented, with noticeable performance degradation when behavior must be inferred from a persona rather than explicitly stated63% similarUnverified同一问题换一种表达方式，模型可能给出截然相反的结论，即提示词敏感性（Prompt Sensitivity）63% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1023600API
curl https://kongchang.com/api/v1/knowledge/claims/1023600MCP
get_claim(id=1023600)