Unverified50% confidenceSolutionExact time
通过记录随机种子可精确复现导致失败的特定顺序，再通过二分法逐步缩小污染源的范围
1
Sources
50%
Confidence
Long-term
Relevance
8/5/2026
First Seen
Sources
Related Claims
Unverified深度学习中的随机种子控制权重初始化、数据加载批次打乱顺序、Dropout神经元随机失活以及优化器中的随机梯度估计等多个随机性来源72% similarUnverified即便固定所有随机种子，在不同硬件或并行度下，训练结果仍可能出现被大模型放大为显著能力差异的细微偏差70% similarUnverified权重初始化的随机性对模型性能影响最深远，数据批次打乱顺序的随机性影响最小68% similarUnverifiedSeed controls randomness at the noise level, and switching the seed with an unchanged prompt only produces minor differences in composition and detail texture, whereas Wildcards introduce an independent random dimension at the semantic level.68% similarUnverifiedIn few-shot prompting, simply changing the order of examples can cause dramatic fluctuations in accuracy, ranging from near-random to near-optimal.67% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/687577API
curl https://kongchang.com/api/v1/knowledge/claims/687577MCP
get_claim(id=687577)