Hierarchical Prompt-Domain Control and Learning for Resource-Constrained Agentic Language Models

2026-05-26

AI Summary

This work addresses the performance degradation of resource-constrained language agents in long-context interactions, caused by prompt-domain invalidation and state drift, alongside challenges in online fine-tuning. The authors propose a hierarchical control-and-learning framework that first distills structured output capabilities into a small model and then employs an Oracle controller to dynamically maintain a valid prompt domain while triggering lightweight online adaptation. The approach formally characterizes prompt-domain feasibility and attention saturation for the first time, decoupling communication-compatible structural learning from task-specific semantic adaptation and emphasizing proactive control over valid prompting states. Experiments on a multi-fidelity Bayesian optimization platform demonstrate that the method significantly outperforms non-hierarchical, distillation-only, and undistilled baselines, achieving a superior trade-off between reliability and cost efficiency.

Abstract

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 correction. We formalize prompt-domain feasibility and attention-induced saturation, motivating control of the effective prompt state rather than reliance on nominal context length. Using Multi-Fidelity Bayesian Optimization as a controlled sequential testbed, we characterize a core deployment failure mode and show improved reliability and cost-efficiency over non-hierarchical, distillation-only, and non-distilled baselines.

Problem

• resource-constrained agentic language models
• prompt-domain feasibility
• structured protocols
• context length limitations
• deployment reliability

Innovation

• hierarchical control
• prompt-domain feasibility
• schema learning
• oracle-supervised fine-tuning
• attention-induced saturation