LLM providers' cost-saving throttling may backfire, research suggests

A new research paper proposes that the common practice of large language model providers degrading service during congestion, such as routing queries to smaller models or truncating context, is economically flawed. The paper argues that this approach incorrectly prices queries based on their arrival rather than the successful delivery of an answer. It introduces models for inference allocation that account for customer lifetime value and the impact of retries, suggesting that throttling can paradoxically increase traffic and permanently degrade service quality under certain conditions. The research also details a method for calculating the 'shadow price of intelligence' to optimize query allocation across different customer classes and times. AI

影响 This research suggests that current LLM provider cost-saving strategies may be counterproductive, potentially leading to increased costs and degraded user experience.

排序理由 The cluster contains a research paper detailing a new model and analysis. [lever_c_demoted from research: ic=1 ai=1.0]

在 arXiv cs.AI 阅读 →

• alphaXiv
• CatalyzeX Code Finder for Papers
• Connected Papers
• DagsHub
• geometric retry multiplier
• Gotit.pub
• Hugging Face
• large language model
• Litmaps
• Newsvendor model
• ScienceCast
• scite Smart Citations
• two-regime transient queue

• 论文
• 基础设施

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