Unverified50% confidenceFactTime unknown
关闭定位注入后，SciMDR模型性能从49.1分骤降至22.8分，降幅达53.6%
1
Sources
50%
Confidence
Medium-term (~90 days)
Relevance
7/2/2026
First Seen
Valid until: 9/30/2026
Sources
SciMDR：7B小模型如何在科研推理上比肩GPT-5
bilibiliAgent创世纪
Related Claims
Unverified将FP32权重降至INT8后，典型Transformer模型基准测试分数下降通常不超过1%，推理速度提升2至4倍，内存占用减少75%66% similarUnverifiedDistilBERT通过蒸馏将BERT参数量减少40%，推理速度提升60%，性能仅损失约3%66% similarUnverified测序成本大约每18个月下降50%，降速超越了半导体行业的摩尔定律64% similarUnverified在仿真中表现优异的策略迁移到真实硬件后，性能通常会下降30%-70%64% similarUnverifiedCodeGraph测试显示Token消耗减少约47%63% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/57292API
curl https://kongchang.com/api/v1/knowledge/claims/57292MCP
get_claim(id=57292)