Hierarchical Prompt-Domain Control and Learning for Resource-Constrained Agentic Language Models
A new arXiv paper (2605.27703) introduces a hierarchical control-and-learning framework for deploying large language models in resource-constrained agentic systems. The method addresses the unreliability of prompt extension in compact models, where growing contexts push models outside their effective prompt domain. The framework first distills a compact model to learn the required output schema, then supervises it online via 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 correction. The paper formalizes prompt-domain feasibility and attention-induced saturation, motivating control of the effective prompt state rather than reliance on nominal context length. Multi-Fidelity Bayesian Optimization is used as a controlled sequential decision-making process. The approach targets scenarios with limited data and compute, aiming to improve reliability and efficiency in agentic deployments.
Enables reliable agentic LLM deployment on resource-constrained devices without costly fine-tuning.