Unverified50% confidenceOpinionExact time
传统评测框架往往只能覆盖预设的测试场景，无法捕捉真实用户交互中的隐性失败
1
Sources
50%
Confidence
Long-term
Relevance
8/27/2026
First Seen
Sources
Related Entities
Related Claims
Unverified针对特定触发词的后门攻击可让模型在大多数情况下表现正常，仅在遇到特定输入时才输出预设错误信息，使常规评测基准难以捕捉74% similarUnverified先前的难负样本挖掘工作大多忽略非干扰项，只关注与查询的相似度，因而更易误选假负样本72% similarUnverifiedThe vast majority of LLM evaluation tools only score answers after they have already been presented to the user, and cannot intercept erroneous outputs in real time.72% similarUnverifiedOnce a platform strengthens parameter validation on the back end, this type of bypass method that constructs non-standard request parameters will immediately become ineffective.70% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/807997API
curl https://kongchang.com/api/v1/knowledge/claims/807997MCP
get_claim(id=807997)