Unverified50% confidenceBenchmarkExact time
使用Reflective Recovery，DeepSeek-R1-Distill-Qwen-7B在AIME 2025上准确率从30.0%提升至37.5%
1
Sources
50%
Confidence
Long-term
Relevance
9/18/2026
First Seen
Sources
Related Entities
Related Claims
Unverified微调后的Qwen3.5-4B模型在BU Bench V1基准上准确率从15%提升至53%68% similarUnverifiedQwen3 27B的IQ4XS量化版本性能相比原始满血版本仅下降约6%63% similarUnverifiedQwen3.6 34B 80层模型在LM Studio上的MTP推测解码接受率约为78%，而原版27B仅约57-58%63% similarUnverified使用Qwen3-27B模型在wikitext-2数据集上以fp32参考logits为基准进行KL散度测量，中位数KLD从0.0023降至0.000001，精度提升约2300倍62% similarUnverifiedQwen2.5-72B 在多项基准测试中接近 Llama-3.1-405B 的表现62% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/931201API
curl https://kongchang.com/api/v1/knowledge/claims/931201MCP
get_claim(id=931201)