Unverified50% confidenceBenchmarkExact time
在SST-2数据集DistilBERT到DistilBERT自蒸馏实验中，可靠性门控方法达到88.800%，RAPID达到88.685%，相比基线87.271%提升约1.4个百分点
1
Sources
50%
Confidence
Medium-term (~90 days)
Relevance
9/11/2026
First Seen
Valid until: 12/10/2026
Sources
Related Entities
Related Claims
Unverified在AG News数据集BERT到DistilBERT蒸馏实验中，可靠性门控关系蒸馏达到94.285±0.054%准确率，RAPID达到94.241±0.025%，均优于交叉熵基线的94.154±0.124%71% similarUnverified探索性能上界实验发现现有验证粒度并非最优，理想情况下计算量（FLOPs）可减少75倍，同时准确率提升1.1%66% similarUnverifiedFastEmbed-rs所使用的量化模型在MTEB基准测试中相比全精度模型的性能下降通常控制在1-2%以内65% similarUnverified在BrowseWebApp Bench基准上，微调后模型的准确率从22%提升至63%64% similarUnverifiedFan团队探索性能上界实验发现，理想情况下计算量FLOPs可减少75倍，同时准确率提升1.1%64% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/895499API
curl https://kongchang.com/api/v1/knowledge/claims/895499MCP
get_claim(id=895499)