Unverified50% confidenceBenchmarkExact time

实验显示低容量会使因果效果降低84%

1

Sources

50%

Confidence

Long-term

Relevance

10/7/2026

First Seen

Sources

拆解语言模型可控性：容量、响应度与对齐的三要素框架

Related Entities

Related Claims

Unverified与全量微调相比，LoRA 微调所需显存通常降低 60%-80%69% similar Unverified实践数据表明经过良好压缩的上下文能将token消耗降低50-80%，并能提升回答准确率67% similar UnverifiedMLA相比GQA可以进一步减少约93%的KV Cache占用67% similar Unverified研究表明，代码评审能减少 60%-90% 的缺陷率66% similar Unverified2023年多项学术研究显示在模拟对抗性环境中主流目标检测算法的准确率可下降40-80%66% similar

Cite This Claim

Stable URI

https://kongchang.com/claim/982943

API

curl https://kongchang.com/api/v1/knowledge/claims/982943

MCP

get_claim(id=982943)