Unverified50% confidenceFactExact time
2018年 Henderson 等人的研究表明许多已发表的 RL 论文中的性能提升在更换随机种子后无法复现
1
Sources
50%
Confidence
Long-term
Relevance
8/26/2026
First Seen
Sources
Related Entities
Related Claims
Unverified2023年多项研究证明，即使经过RLHF对齐的模型也难以完全抵抗精心设计的间接提示词注入65% similarUnverified在某些任务上，仅因随机种子不同导致的性能波动幅度可能超过某些论文声称的方法创新带来的提升63% similarUnverified低精度引入的数值扰动会在RL的多轮迭代中通过策略梯度的乘积形式被放大，可能导致训练崩溃或模型性能退化60% similarUnverifiedA 2018 study by the Google Brain team showed that simply changing the random seed could cause performance fluctuations of several percentage points in certain models.58% similarUnverified2019年Dodge等人的研究发现NLP领域大量论文缺乏足够的实现细节，导致独立研究者无法复现原始结果57% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/805337API
curl https://kongchang.com/api/v1/knowledge/claims/805337MCP
get_claim(id=805337)