Unverified50% confidenceBenchmarkExact time
在RULER-16K数据集的等预算对照实验中，KVFetch使逐字拷贝得分从0.8恢复到78.4
1
Sources
50%
Confidence
Long-term
Relevance
10/8/2026
First Seen
Sources
Related Entities
Related Claims
Unverified在RULER-16K实验中，KVFetch使13项任务的平均分提升了+8.4，增益主要集中在需要顺序访问的任务上65% similarUnverified经验内化微调把Qwen3-8B-27B的一次性得分从37.07提升到48.22，同时生成Token从641万降到57万（减少11倍）59% similarUnverified使用Reflective Recovery，DeepSeek-R1-Distill-Qwen-7B在AIME 2025上准确率从30.0%提升至37.5%59% similarUnverified该量化模型相较于BF16全精度版本保留了约77%的准确率58% similarUnverified在分歧富集数据集上，gpt-4o-mini裁判与两位作者共同确认的黄金标准之间的Cohen's kappa一致性系数仅为0.0458% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/986758API
curl https://kongchang.com/api/v1/knowledge/claims/986758MCP
get_claim(id=986758)