Paper Overview
- Field: LLM
- Authors: Joan Vendrell Gallart, Russell Bent, Michael Grosskopf
- Published: 2026-05-28
- arXiv: 2605.27703
- Formalizes prompt-domain feasibility and attention-induced saturation, arguing that systems should control the effective prompt state rather than rely on nominal context length.
- Uses multi-fidelity Bayesian optimization as a controlled sequential testbed to characterize core deployment failure modes.
- Demonstrates improved reliability and cost efficiency compared to non-hierarchical, distillation-only, and non-distillation baselines.
Summary
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.
The paper proposes a hierarchical control-and-learning framework:
1. Schema distillation: A compact model is first distilled to learn the required output schema. 2. Oracle-controller loop: The model is then supervised online. The controller monitors protocol validity and semantic performance. 3. Prompt-domain projection: Accumulated histories are projected into a feasible prompt domain. 4. Drift-triggered fine-tuning: Lightweight oracle-supervised fine-tuning is triggered when drift is detected.
This separates schema learning for communication compatibility from semantic adaptation for task-level correction.