[论文] Hierarchical Prompt-Domain Control and Learning for Resource-Cons...

论文概要

研究领域: LLM
作者: Joan Vendrell Gallart, Russell Bent, Michael Grosskopf
发布时间: 2026-05-28
arXiv: 2605.27703

中文摘要

大语言模型越来越多地部署在智能体系统中,需要遵循结构化协议、适应演进状态,并在内存、延迟和成本约束下运行。在此类场景中,提示扩展并不可靠:增长的上下文可能将紧凑模型推出其有效提示域,而部署时微调受限于稀缺数据和计算资源。本文提出了一种分层控制与学习框架:紧凑模型首先通过蒸馏学习所需输出模式,然后由预言机-控制器循环在线监督。控制器监控协议有效性和语义性能,将累积历史投影到可行的提示域,并在漂移时触发轻量级预言机监督微调。这分离了用于通信兼容性的模式学习与用于任务级纠正的语义适应。研究形式化了提示域可行性和注意力诱导饱和,论证了控制有效提示状态而非依赖名义上下文长度。使用多保真贝叶斯优化作为受控顺序测试平台,表征了核心部署失败模式,并展示了相比非分层、纯蒸馏和非蒸馏基线的可靠性和成本效率提升。

原文摘要

Large Language Models are increasingly deployed inside agentic systems, where they must follow structured protocols, adapt to evolving states, and operate under memory, latency, and cost constraints. In such regimes, prompt extension is unreliable: growing contexts can push compact models outside their effective prompt domain, while deployment-time fine-tuning remains limited by scarce data and compute. We propose a hierarchical control-and-learning framework in which a compact model is first distilled to learn the required output schema, then supervised online by an oracle-controller loop. The controller monitors protocol validity and semantic performance, projects accumulated histories into a feasible prompt domain, and triggers lightweight oracle-supervised fine-tuning under drift. This separates schema learning for communication compatibility from semantic adaptation for task-level c...

自动采集于 2026-05-29

#论文 #arXiv #LLM #小凯