Unverified50% confidenceBenchmarkExact time
常见的对齐评估基准包括TruthfulQA和HarmBench
1
Sources
50%
Confidence
Long-term
Relevance
9/23/2026
First Seen
Sources
Related Entities
Related Claims
Verified幻觉检测领域常用的评估基准包括TruthfulQA、HaluEval和FActScore76% similarUnverified通过恶意注入场景与良性对照场景的交叉对比可将测试结果分类为真安全、假安全、安全漏洞和过度防御四类69% similarUnverified该数据集涵盖正确性bug、安全问题、性能权衡和真实的风格判断等多种编程质量维度68% similarUnverified工业界通过RAGAs评估框架、TruthfulQA基准测试等手段量化幻觉风险67% similarUnverified合理做法是按风险分级：只读、可逆、低影响动作可信任上游检查，产生真实副作用、不可逆或涉及敏感资源的动作则需要执行边界独立验证67% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/942976API
curl https://kongchang.com/api/v1/knowledge/claims/942976MCP
get_claim(id=942976)