LLM-Derived Prompts (LDPs)

• LLM-Derived Prompts (LDPs) are prompts generated through LLM introspection, automated search, and optimization, enabling adaptive and personalized prompt engineering.
• They integrate techniques like self-calibration, automated segmentation, and discrete optimization to improve evaluation, adversarial red-teaming, and user modeling.
• Empirical results show LDPs boost accuracy, consistency, and calibration metrics in domains such as legal QA, multi-agent dialogue, and traceability tasks.

LLM-Derived Prompts (LDPs) constitute a class of prompts autonomously constructed, parameterized, or optimized by LLMs, often through direct model introspection, automated search, or data-driven adaptation. Contrary to static, manually crafted prompts—commonly termed expert-designed prompts—LDPs exploit the LLM’s own latent biases, contextual awareness, and generative capacity, thereby enabling personalized, adaptive, or adversarial workflow integration. Recent research has established LDPs as a foundational technique for prompt engineering across domains including evaluative reasoning, adversarial red-teaming, multi-agent dialogue, automated traceability, and user behavior modeling.

1. Formal Definition and Theoretical Foundations

An LLM-Derived Prompt is any prompt $P^*$ produced by an LLM, either by explicit introspection (e.g., self-generated scales, automated segmentations), discrete optimization, or by sampling conditioned on target outputs. Let $\theta$ denote the (fixed) parameters of an LLM, and $p_\theta(y|x, P)$ the distribution over outputs $y$ given input $x$ and prompt context $P$. The prompt-bias functional

$$\mathrm{IB}_\theta(P) := \mathbb{E}_{x\sim\mathcal{D}}\left[ H\big(p_\theta(\cdot|x,P)\big) \right]$$

(where $H$ is Shannon entropy) captures the degree of alignment between prompt $P$ and the LLM’s inductive bias. Small prompt rewordings can yield substantial shifts in $p_\theta$, exposing prompt sensitivity. The Inductive Bias Extraction and Matching (IBEM) strategy operationalizes this by extracting model-preferred artifacts (e.g., Likert scales or reasoning patterns) and folding them back into future prompts, empirically reducing prediction uncertainty and improving calibration (Angel et al., 14 Aug 2025).

2. Construction Methodologies

LDPs are constructed through a diversity of techniques:

• Self-Calibration via Introspection: The IBEM strategy automatically generates model-preferred rating scales or reasoning schemas for each sub-metric. These are then reincorporated as part of downstream prompts, tightly matching the LLM’s own bias. Pseudocode for IBEM involves prompt-based extraction, model-guided candidate rating, score aggregation, and prediction selection (Angel et al., 14 Aug 2025).
• Automated Segmentation: For evaluation tasks, such as legal text judgment, the model segments long-form outputs into atomic, verifiable assertions—Legal Data Points (LDPs)—which are then tagged for correctness or related error types. Extraction employs a single-pass prompt plus post-processing, producing concise, graded units for reference-free metrics (Enguehard et al., 8 Oct 2025).
• Discrete or Continuous Optimization: In adversarial and personalization contexts, LDPs can be viewed as vectors or token sequences found by optimizing an objective (e.g., reward or alignment function). This encompasses alternating optimization-fine-tuning cycles (as in AutoPrompT), graph-based embeddings mapped to personalized soft prompts for sequential models (Liu et al., 28 Oct 2025, Song et al., 2024), and conditional sampling under joint prompt–response distributions (diffusion LLMs) (Lüdke et al., 31 Oct 2025).
• Discriminative and Policy-Parameterized Control: Prompt templates may be parameterized by explicit state components (retrieved memory, persona, knowledge) and dynamic weights (policy-parameterized prompts for multi-agent dialogue), or selected via discriminative scoring over candidate generations (direct/inverse/hybrid prompts for self-improving output reliability) (Ahn et al., 2024, Bo et al., 10 Mar 2026).

3. Evaluation, Performance, and Empirical Results

LDPs have demonstrated significant empirical advantages across multiple application domains. Key findings include:

Task/Domain	Baseline Metric	LDP Metric	Relative Gain	Source
WikiHow Ranking (5-way)	Accuracy 48.1%	59.9%	+24.5%	(Angel et al., 14 Aug 2025)
Legal QA (LegalBench)	IAA 0.77	0.88	+11% inter-annotator agreement	(Enguehard et al., 8 Oct 2025)
T2I Red-teaming (SLD-MAX)	RSR 62.5% (P4D-Union)	70.5% (APT)	+13% Red-teaming Success Rate	(Liu et al., 28 Oct 2025)
Smart Space Prediction	Macro-F1 0.691	0.727	+3.6 pp	(Song et al., 2024)

In traceability, chain-of-thought (CoT) augmented LDPs substantially improved recall and precision (e.g., recall 46% $\theta$0 92%, precision 18% $\theta$1 37.9% on CM1) (Rodriguez et al., 2023). For classification and ranking, IBEM yields up to 27% improvements in mean reciprocal rank and up to 19.5% in Macro F1 (Angel et al., 14 Aug 2025). In adversarial prompting, diffusion-based LLMs amortize search, achieving attack success rates of up to 100% on open-source models and 53% on ChatGPT-5, with low perplexity and efficient compute utilization (Lüdke et al., 31 Oct 2025).

4. Application Domains and Architectures

• Evaluation and Fact Segmentation: LDPs enable reference-free, fine-grained grading of legal or fact-based outputs. Atomic assertions segmented as LDPs are independently scored, allowing for granular precision, recall, and omission analysis (Enguehard et al., 8 Oct 2025).
• Adversarial Red-Teaming: LLMs auto-generate adversarial suffixes or full prompts targeting output alignment vulnerabilities. Dual constraints—perplexity (human readability) and banned-token filtering—enable robust filter evasion and high transferability across T2I and text models (Liu et al., 28 Oct 2025, Lüdke et al., 31 Oct 2025).
• Personalized User Modeling: Graph-to-prompt embeddings (soft continuous prompts) derived from user behavior histories inform individualized predictions in smart environments, especially enhancing generalization on sparse and rare events (Song et al., 2024).
• Multi-Agent Control: Policy-parameterized prompts govern dialogue behavior in LLM-based agent simulations, supporting structured negotiation, evidence use, stance maintenance, and adaptive conversational policy without further LLM training (Bo et al., 10 Mar 2026).
• Self-Improving Generation: Discriminative prompt schemes (direct, inverse, hybrid) exploit the LLM’s own uncertainty and candidate ranking capacities to select more accurate or reliable generations, validated across math benchmarks (Ahn et al., 2024).

5. Limitations, Failure Modes, and Best Practices

Several limitations are identified:

• Prompt Sensitivity and Model Dependency: LDP effectiveness depends on the LLM’s intrinsic capacity to introspect, generate meaningful artifacts, and maintain calibration. Quality and consistency may degrade for weaker LMs or under poorly specified metrics (Angel et al., 14 Aug 2025).
• Variance Induced by Stochastic Scale Extraction: Repeated LDP extraction (e.g., through IBEM) introduces trial-to-trial variability (±2–3pp accuracy). Certain distractor types may cause more ties or instabilities (Angel et al., 14 Aug 2025).
• Inference Overhead: For fine-grained or multi-metric extraction, LDPs may require multiple LLM calls per sample, raising compute cost (Angel et al., 14 Aug 2025).
• Domain-Specific Tuning: Calibration (e.g., granularity in legal LDP segmentation or balance of omitted vs. hallucinated assertions) must be tuned for target applications (Enguehard et al., 8 Oct 2025).
• Reference-Free Limitations: In legal evaluation, subjectivity in “relevance” implies modest improvements in inter-annotator agreement relative to correctness; interpretability and traceability of disagreements are improved, however (Enguehard et al., 8 Oct 2025).
• Adversarial Generalization: Adversarial LDPs (diffusion, red-teaming) may still transfer to robustly tuned models, but sample efficiency and harmfulness can vary with the fidelity of joint prompt–response distribution approximation and thresholded candidate selection (Lüdke et al., 31 Oct 2025).

Best practices include: iterative prompt refinement (“prompt surgery” using model feedback), explicit sub-metric extraction, chaining reasoning steps, pre-ranking candidates, and leveraging contextual or persona embeddings in multi-agent scenarios (Rodriguez et al., 2023, Bo et al., 10 Mar 2026).

6. Future Directions and Research Frontiers

Prominent avenues identified include:

• Automated Sub-Metric Discovery: Leveraging information-theoretic approaches to automatically select sub-metrics for LDP extraction (e.g., mutual information with target labels) (Angel et al., 14 Aug 2025).
• Hybridization with Light RL: Learning optimal prompt parameter schedules (e.g., rules, weights) for policy-parameterized LDPs through differentiable or bandit approaches (Bo et al., 10 Mar 2026).
• Extending to Generative and Open-Ended Tasks: Adapting LDP analytical patterns—such as chain-of-thought pattern extraction or segment-based scoring—to summarization, translation, or dialogue (Angel et al., 14 Aug 2025).
• Distillation and Efficient Adaptation: Reducing inference cost by distilling multi-call LDPs into single static prompts, or using lightweight prompt-adapter networks (Angel et al., 14 Aug 2025).
• Test-Time Personalization: Dynamic adaptation of prompts to user histories, context drift, or multi-modal (e.g., audio/video) signals to support long-term deployment in interactive and personalized systems (Song et al., 2024).
• Standardized Evaluation Frameworks: Developing unified, model-agnostic metrics for evaluating LDP efficacy and supporting reproducible research across legal, adversarial, and open-domain tasks (Enguehard et al., 8 Oct 2025).

LDPs recast prompt engineering as an LLM-centric, data-driven paradigm, systematically leveraging the model’s own reasoning, preference, and generative capabilities to optimize or personalize downstream task performance, interpretability, and robustness.