CPGPrompt: Unified Prompt Engineering

• CPGPrompt is a framework that integrates prompt engineering, discrete search, and continuous learning to drive LLM-based decision support across domains.
• It leverages structured decision trees and LLM-guided pipelines to translate narrative clinical guidelines into interpretable, auditable decision processes.
• Empirical evaluations reveal trade-offs in recall and precision, highlighting the need for quantitative criteria and manual validation in complex applications.

CPGPrompt

CPGPrompt encompasses a set of algorithmic paradigms and frameworks that unify prompt engineering, discrete prompt search, and continuous prompt learning for large models across diverse domains—ranging from code generation to continual learning, graph adaptation, protein conformation modeling, and clinical decision support. The term "CPGPrompt" refers to (1) specialized algorithms for constructing decision-structural prompts (notably in medical guideline automation), (2) methods leveraging structured, graph-based, or soft prompts to condition frozen backbone models, and (3) the broader class of prompt optimization frameworks that operate without LLM fine-tuning. This article surveys key instantiations, mathematical principles, and empirical results associated with CPGPrompt approaches, with a focus on methods and systems introduced or analyzed under this name or exact acronym.

1. Translating Narrative Guidelines via CPGPrompt Decision Trees

At the core of the original CPGPrompt system is the translation of long-form clinical practice guidelines into machine-actionable, LLM-executable decision trees (Deng et al., 7 Jan 2026). The framework consists of a semi-automatic pipeline with the following stages:

1. Segmentation: The source guideline is partitioned into semantically coherent segments (e.g., diagnostic criteria, referrals).
2. LLM-Assisted Subtree Generation: Each segment is subjected to a specialized prompt ("Extract from this passage all decision nodes... output JSON...") to induce a subtree with structured nodes (name, type, criteria, thresholds, actions, and yes/no branches).
3. Synthesis and Merging: Subtrees are consolidated into a global decision tree $T = (N, E)$ with explicit edge connections representing clinical choices.
4. Manual Clinical Validation: Clinical experts review and edit misparsed or misaligned nodes, ensuring fidelity to underlying medical logic.
5. Tree-Based LLM Query Execution: For each patient vignette, the LLM system navigates the decision tree, querying at each node via templated yes/no or multi-criteria prompts, with path traversal yielding guideline-backed decisions.

The tree is formally defined as a tuple $T = (N, E)$, where each node $n$ encodes node type, criteria $C$, threshold $\theta$ (for multi-criteria), action $A$ (for terminal nodes), and branching structure. The traversal uses a strict prompting protocol to collect model answers and maintain an auditable log at each decision step.

2. Auto-Prompting Pipeline and LLM Integration

The CPGPrompt platform operationalizes guideline logic via an LLM-based execution engine that accepts a patient vignette as context and traverses the structured decision tree, dynamically selecting child nodes according to model predictions and criteria satisfaction (Deng et al., 7 Jan 2026). This process is implemented as follows:

• Node Query Construction: For "simple" nodes, the system issues prompts such as "Does the patient have {CRITERION}? Answer (Yes/No):". For "multi" nodes, each criterion is separately queried; the results are aggregated and compared to the node's threshold to determine branching.
• Prompt Template: The prompt instantiates slot variables (VIGNETTE, FEATURE_NAME, etc.) and leverages explicit answer format constraints to minimize ambiguity.
• Audit and Logging: All prompt/response pairs are stored, enabling full auditability and reproducibility of inference pathways.

This architecture ensures that even complex, nested, or longitudinal clinical logics can be executed over free-text patient notes with precise adherence to the original guidelines, albeit subject to LLM performance on nuanced criteria.

3. Synthetic Vignette Generation, Evaluation Protocol, and Empirical Performance

To benchmark CPGPrompt's decision correctness and robustness, synthetic patient vignettes are systematically generated:

• Category Stratification: Vignettes encompass single-criteria, multi-criteria, contrastive-criteria (negative and positive statements), and exclusion-criteria categories.
• Controlled Sampling: For each leaf guideline action, vignettes are synthesized to match positive and negative paths, with clinician review of representative subsets.
• Evaluation Metrics: Binary referral and multi-class pathway assignments are measured using standard precision, recall, F1, and macro-averaged F1 across categories.

Results (all domains, five runs):

Domain	Binary F1	Multi-class F1
Headache	0.90 ± 0.00	0.44 ± 0.01
Low Back Pain	1.00 ± 0.00	0.72 ± 0.03
Prostate Cancer	1.00 ± 0.00	0.77 ± 0.04

Exclusion-criteria multi-class F1 was low for headache (0.06) due to over-referral when negation is challenging for the LLM. Higher F1s are observed where criteria (e.g., laboratory values for prostate cancer) are quantitative and explicit.

4. Domain-Specific Challenges and Error Modes

Performance analysis by domain (Deng et al., 7 Jan 2026):

• Headache: Lowest multi-class F1, over-referral in exclusion cases due to negation errors ("denies neurological deficits"), defaulting to safety-first referral.
• Lower Back Pain: Challenges with temporal reasoning and duration parsing ("not improving after 1–6 weeks"); noisy narrative degrades assignment precision, especially in longer vignettes.
• Prostate Cancer: Highest robustness, attributed to clear numeric criteria (e.g., PSA thresholds), achieving F1 = 1.00 in exclusion cases; integrating borderline cases with multiple factors remains an edge case.

These outcomes underscore the dependence of prompt-executed decision support on the original guideline's formality and on the ability of LLMs to parse structured versus subjective/narrative features.

5. Theoretical Properties, Limitations, and Generalization

Interpretability and Auditability: Each LLM decision is logged with the triggering prompt and model answer, enabling full transparency; the use of a structured decision tree enforces non-hallucinatory reasoning (Deng et al., 7 Jan 2026).

Robustness and Over-Referral Tradeoff: The system achieves high recall (often 1.00) but may sacrifice precision in scenarios involving subjective or heavily negated criteria, aligning with safety-first decision-making but posing burdens in terms of false positives.

Stated Limitations:

• Efficacy is limited by the model's performance on narrative negation and temporality.
• Synthetic rather than real EHR text is used for benchmarking.
• The JSON schema and execution model are not designed natively for longitudinal data or repeated encounters.
• Manual curation is required to correct errors in criteria extraction and segmentation.

Proposed Enhancements:

• Incorporation of chain-of-thought or temporal reasoning modules to handle sequential and temporal logic.
• Human-in-the-loop validation for subjective or ambiguous features.
• Automation of flowchart-to-graph transformations and support for updateable, longitudinal decision trees.

6. Recommendations and Broader Applicability

Recommendations for extending or deploying CPGPrompt-like systems (Deng et al., 7 Jan 2026):

• Favor initial application to domains with rule-based, quantitative, or laboratory-test-driven guidelines.
• Invest in manual validation of multi-criteria thresholds and time frames during tree construction.
• Use tightly structured or bullet-pointed input notes for inference to mitigate narrative ambiguity.
• Pilot on synthetic or semi-synthetic test data and incorporate real-world clinician oversight during staged rollout.

While first instantiated in clinical guideline translation, the general CPGPrompt paradigm—structuring high-level objectives or logic as a prompt-executed tree—has clear analogs in algorithmic code generation (e.g., Chain of Grounded Objectives (Yeo et al., 23 Jan 2025)), continual learning (prompt-based and graph-prompt frameworks (Wang et al., 10 Feb 2025, Gao et al., 2024, Huang et al., 26 Sep 2025, Peng et al., 2023)), and protein or multimodal modeling (schema-constrained or graph-injected prompts (Zhang et al., 2022, Peng et al., 2023)). All these share the central feature of explicit structure or semantics infused into frozen or plug-and-play model architectures purely via prompt configuration.

---

References:

• "CPGPrompt: Translating Clinical Guidelines into LLM-Executable Decision Support" (Deng et al., 7 Jan 2026)
• "Chain of Grounded Objectives: Bridging Process and Goal-oriented Prompting for Code Generation" (Yeo et al., 23 Jan 2025)
• "Prompt-Driven Continual Graph Learning" (Wang et al., 10 Feb 2025)
• "Consistent Prompting for Rehearsal-Free Continual Learning" (Gao et al., 2024)
• "One Prompt Fits All: Universal Graph Adaptation for Pretrained Models" (Huang et al., 26 Sep 2025)
• "MMGPL: Multimodal Medical Data Analysis with Graph Prompt Learning" (Peng et al., 2023)
• "Prompt-Guided Injection of Conformation to Pre-trained Protein Model" (Zhang et al., 2022)