A new paper proposes a theoretical framework that unifies economic principles of marginal utility with machine learning concepts like matrix factorization and the Key-Value (KV) cache in transformer language models. This framework suggests an allocation rule for retaining top dimensions based on their eigenvalue exceeding a shadow price. The paper applies this to geo-mining document extraction, detailing a multi-pass inference protocol and a layer-wise TIES model merging procedure. Empirically, an 11.2-million-parameter classifier achieved 90.0% accuracy on a uranium-exploration corpus with low latency and cost, while a diagnostic revealed and corrected a degenerate merging mode. AI
影响 Introduces a novel theoretical lens for optimizing LLM inference efficiency and model merging techniques.
排序理由 The cluster contains a single arXiv paper detailing theoretical and empirical contributions to AI model inference and optimization. [lever_c_demoted from research: ic=1 ai=1.0]
- alphaXiv
- Caroline Gans Combe
- CatalyzeX
- Conditional Value-at-Risk
- DagsHub
- Key-Value (KV) cache
- LoRA+
- TIES model
- transformer language models
AI 生成摘要 · Google Gemini · 来自 1 个来源。 我们如何撰写摘要 →
