PulseAugur
实时 06:57:15
English(EN) Temperature 0 is not reproducible. I measured 30 percent of my output changing between identical runs.

LLM推理批处理导致显著的输出不可复现性

大型语言模型(LLM)即使在温度设置为0时也会表现出不可复现的输出,这是由于推理过程中的批处理。这是因为批处理计算涉及与非批处理不同的矩阵乘法和浮点运算,导致logits的微小差异,当概率接近时会改变token的选择。作者在A/B测试比较过滤器的有效性时发现了一个不可能的结果,从而发现了这个问题,揭示了底层模型输出在运行之间存在显著差异(约30%的记录发生变化),这并非由于采样,而是由于批处理。 AI

影响 LLM输出的可复现性对于研究和生产至关重要;这一发现表明批处理会引入微妙但显著的变化,影响基准测试和A/B测试。

排序理由 该条目详细介绍了关于LLM推理行为的技术发现。[lever_c_demoted from research: ic=1 ai=1.0]

在 dev.to — LLM tag 阅读 →

AI 生成摘要 · Google Gemini · 来自 1 个来源。 我们如何撰写摘要 →

LLM推理批处理导致显著的输出不可复现性

本文如何被排名

Signal score
28 / 100
Composite score across the factors below. Higher = stronger signal that this story matters right now.
Newsworthiness bucket
Tool
该条目详细介绍了关于LLM推理行为的技术发现。[lever_c_demoted from research: ic=1 ai=1.0]
Source corroboration
Single-source cluster
Only one publisher covered this so far. Single-source stories can still rank when the publisher is high-authority, but they lack cross-source corroboration.
Topics
infra
Editorial topic classification. Feeds into how the story surfaces on /topic/<slug> hub pages and into the per-entity coverage mix.
AI-industry relevance
High
Clearly on-topic for AI-industry coverage.
Story freshness
Breaking (< 6h)
Fresh story with cross-source coverage still developing. Ranking may shift as more sources report.

完整方法见我们的编辑标准

报道来源 [1]

  1. dev.to — LLM tag TIER_1 English(EN) · Rickesh T N ·

    Temperature 0 不可复现。我测量到 30% 的输出在相同运行之间发生变化。

    <p>I set temperature to 0, ran the same document through the same model twice with<br /> the same prompt, and got different answers. Not subtly different. Roughly a third<br /> of the extracted records changed.</p> <p>The cause is not sampling. It is batching.</p> <div class="tab…