Unverified50% confidenceFactExact time
99.3%的通过率是在可重建、可评分的子集上得出的，而非全部100个版本，其中V1–V4无法恢复被排除，V6是早期不稳定基线无法精确聚合
1
Sources
50%
Confidence
Long-term
Relevance
8/24/2026
First Seen
Sources
Related Entities
Related Claims
Unverified使用100个以上主成分时，重建结果与原图几乎无法区分62% similarUnverified作者强调99.3%不是CFC准确率的普遍值，也不是Gemini的通用可靠性评分，只是该特定可复现决策闭合基准上的分数62% similarUnverified在漏洞检测基准测试中，让最新模型运行五次，只有50%的漏洞被稳定发现，相比确定性检查工具LLM只发现75%的问题，F1分数仅为40%61% similarUnverifiedINT8 quantization loses approximately 0.5% accuracy while halving VRAM usage; INT4 quantization compresses VRAM to 1/4 of the original with an accuracy loss of approximately 1–3%.60% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/792896API
curl https://kongchang.com/api/v1/knowledge/claims/792896MCP
get_claim(id=792896)