Hierarchical Prompt-Domain Control for Resource-Constrained Agentic Language Models
A new framework for controlling and learning hierarchical prompt domains in resource-constrained agentic language models is proposed. The approach involves distilling a compact model to learn output schemas, then supervising it online via an oracle-controller loop that monitors protocol validity and semantic performance. The controller projects histories into feasible prompt domains and triggers lightweight fine-tuning under drift, separating schema learning from semantic adaptation. This addresses challenges of prompt extension unreliability and limited deployment-time fine-tuning data and compute.
Key facts
- arXiv:2605.27703v1
- Announce Type: new
- Large Language Models are increasingly deployed in agentic systems
- Must follow structured protocols and adapt to evolving states
- Operate under memory, latency, and cost constraints
- Prompt extension is unreliable in compact models
- Deployment-time fine-tuning limited by scarce data and compute
- Proposed framework: hierarchical control-and-learning
- Compact model first distilled to learn output schema
- Supervised online by an oracle-controller loop
- Controller monitors protocol validity and semantic performance
- Projects accumulated histories into feasible prompt domain
- Triggers lightweight oracle-supervised fine-tuning under drift
- Separates schema learning from semantic adaptation
Entities
—