Unverified50% confidenceOpinionExact time
The LLM-judge approach may give inconsistent conclusions for the same suspicious text when the temperature parameter is non-zero, and attackers can wait for passage by repeatedly retrying.
1
Sources
50%
Confidence
Long-term
Relevance
10/13/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedLLM在不同温度参数、上下文长度、随机种子下可能对同一攻击payload给出截然不同的响应，因此LLM安全测试的通过标准本质上是统计性的74% similarUnverifiedLLM judge results are affected by factors such as temperature, prompt wording, and model version updates, causing the same set of test cases to potentially produce different conclusions when run at different times.73% similarUnverifiedSummaries generated by LLMs are probabilistic; the same input may produce different results at different temperature settings, posing a reproducibility risk.73% similarUnverified即便temperature设为0，LLM在长对话上下文中仍可能出现幻觉或遗漏必要披露，不宜依赖模型生成强监管行业的合规话术73% similarUnverified业界应对大模型非确定性的测试策略包括将温度设为 0、使用 LLM-as-Judge，以及使用 Mock 数据隔离 API 调用68% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1020192API
curl https://kongchang.com/api/v1/knowledge/claims/1020192MCP
get_claim(id=1020192)