Unverified50% confidenceFactExact time
训练-部署鸿沟指模型因训练与部署环境的数据分布偏移、软件依赖版本差异或硬件指令集不同而导致的性能下降
1
Sources
50%
Confidence
Long-term
Relevance
7/17/2026
First Seen
Sources
Related Claims
Unverified参数量相似的情况下，训练数据质量分布的差异往往比架构差异更能决定特定场景的表现上限78% similarUnverifiedBehavioral cloning suffers from the distribution shift problem: during training the policy only encounters the state distribution covered by expert trajectories, and during deployment the accumulation of small deviations causes the actual state distribution visited to gradually diverge from the training distribution.76% similarUnverified不同模型系列的工具调用输出格式差异根源于指令微调阶段使用的训练数据差异76% similarUnverified针对特定工具做专项强化训练的副作用是对非训练分布工具的泛化能力下降，学术上称为分布外泛化失败76% similarUnverified域偏移是指训练数据分布与实际部署场景数据分布之间的系统性差异，是迁移学习和AI工程落地的核心挑战之一74% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/541096API
curl https://kongchang.com/api/v1/knowledge/claims/541096MCP
get_claim(id=541096)