prompt engineering

Multilingual large language models (LLMs) exhibit poor generalization on low-resource languages and heavily rely on parameter-intensive fine-tuning. Method: We systematically review 36 papers (2021–2023), covering 250 languages, 30 NLP tasks, and 39 prompting techniques, and propose the first multidimensional classification and analytical framework integrating language families and resource levels (high/low). We introduce model-agnostic prompting strategies—including natural-language prompt design, zero-/few-shot cross-lingual transfer, knowledge elicitation, and templating—empirically validated on mT5, XGLM, and LLaMA-2-Multilingual. Results: The synthesized state-of-the-art prompting strategies yield an average performance gain of 12.7% on low-resource language tasks without any parameter updates. Our core contribution is the establishment of the first interpretable, transferable theoretical framework and practical guide for multilingual prompt engineering.

Manual prompt engineering for large language models (LLMs) is labor-intensive, empirically driven, and lacks theoretical foundations. Method: This paper pioneers modeling prompt optimization as a linear feedback control system—using the deviation between target and actual outputs as the error signal, and dynamically updating prompts via proportional (P), integral (I), or proportional-integral-derivative (PID) controllers. Contribution/Results: It establishes the first rigorous mapping between LLM prompt optimization and classical control theory, yielding analytical tractability, tunable parameters, and theoretical interpretability. Experiments across diverse tasks demonstrate that closed-loop prompt optimization significantly improves convergence speed and stability: iteration counts decrease by 62% compared to manual tuning, while exhibiting strong robustness to controller parameter variations. This framework overcomes the fundamental bottleneck in systematically optimizing nonlinear, black-box LLMs.

Prompt engineering for large language models (LLMs) faces key bottlenecks: heavy reliance on manual design, static updates, coarse-grained editing, and poor reusability of empirical knowledge. To address these, this paper proposes PromptFlow—a modular, TensorFlow-inspired framework for end-to-end prompt optimization. Methodologically, it introduces (1) differentiable, fine-grained prompt editing; (2) a hybrid optimization strategy integrating gradient-based meta-learning and reinforcement learning to enable dynamic policy selection and cross-task LLM experience transfer; and (3) a unified computational graph comprising meta-prompts, prompt operators, optimization modules, and evaluation components. Empirically validated across diverse NLP tasks, PromptFlow achieves substantial downstream performance gains using only minimal labeled data—demonstrating clear superiority over conventional static prompt engineering approaches.

This work addresses the challenge that users struggle to precisely control subjective preferences—such as tone and style—in large language model (LLM) generation through natural language prompts alone. To overcome this limitation, the authors propose Malleable Prompting, a novel approach that automatically parses natural language expressions of preference and maps them to intuitive graphical user interface (GUI) controls, such as sliders and dropdown menus, augmented with real-time visual feedback. During decoding, the method dynamically modulates the token probability distribution to enable transparent and fine-grained control over generation outcomes. User studies demonstrate that, compared to conventional textual prompting, Malleable Prompting significantly improves alignment with user preferences and is consistently perceived as more controllable and interpretable.

Current prompt engineering methodologies overemphasize automation techniques (e.g., role-playing, chain-of-thought) while neglecting users’ ability to articulate clear, customized requirements—resulting in low-quality prompts for complex tasks. Method: This paper introduces Requirement-Oriented Prompt Engineering (ROPE), a novel paradigm centered on *requirement quality* as the core training objective. ROPE establishes a human-centered instructional framework integrating expert annotation, structured training tasks, and LLM-driven real-time feedback to iteratively refine requirement formulation. Contribution/Results: Empirical analysis confirms a strong positive correlation between input requirement quality and downstream LLM performance. A randomized controlled trial with 30 novices demonstrates that ROPE improves task success rate by 20%—significantly outperforming conventional prompt training (+1%)—and this gain is not replicable via automated prompt optimization alone. The framework yields a scalable, pedagogically grounded teaching toolkit for effective prompt authoring.

Existing prompt optimization methods rely on manual design or supervised training, suffering from poor generalizability and strong task coupling. To address this, we propose the Hierarchical Multi-Agent Workflow (HMAW), a zero-shot, task-agnostic, human-free, and model-training-free framework for autonomous prompt generation. HMAW employs coordinated large language model (LLM) agents—each assuming distinct roles—to jointly perform instruction refinement, semantic alignment, and dynamic feedback, enabling end-to-end high-quality prompt construction without external annotations or domain-specific priors. Evaluated across multiple benchmarks, HMAW significantly improves LLM answer accuracy while generating more comprehensive and context-adaptive prompts. This work establishes the first fully autonomous, hierarchically collaborative zero-shot prompt optimization paradigm, overcoming fundamental limitations of conventional manual prompt engineering and supervised fine-tuning approaches.

This work addresses the limitations of large language models in practical deployment, where textual prompts often fail to enable efficient, stable, and inference-only customization. To overcome this, the paper proposes opening vector prompts as a standardized user interface, establishing a novel customization paradigm. Through vector prompt tuning, attention mechanism analysis, and security evaluation under black-box threat models, experiments demonstrate that vector prompts consistently improve performance with enhanced supervision signals, whereas textual prompts saturate early. Moreover, vector prompts induce globally dense attention patterns, revealing superior controllability and greater potential for model customization compared to conventional textual prompting.

In knowledge-intensive tasks, non-expert users struggle to craft effective prompts, limiting the practical efficacy of large language models (LLMs). To address this, we introduce PromptGuide—the first interactive, human-in-the-loop prompt engineering system, grounded in four empirically derived design principles. PromptGuide integrates LLM-based reasoning with principled human-computer interaction techniques to form a prompt enhancement framework that provides real-time feedback, strategy recommendations, and closed-loop optimization of user intent. A double-blind randomized controlled study (N=80) demonstrates that users employing PromptGuide achieve a significantly higher median task accuracy (78.3 vs. 61.7 for controls; p<0.01), alongside substantial improvements in efficiency, usability, and operational autonomy (all p<0.01). This work establishes a validated, interactive paradigm for prompt engineering that lowers the barrier to LLM adoption and strengthens user agency.

This work addresses the limited interpretability and poor transferability of existing soft prompt tuning methods, particularly when applied to closed-source large language models. The authors propose the first end-to-end framework for translating optimized soft prompts into human-readable natural language (hard) prompts, leveraging a dedicated model trained across multiple datasets. Evaluated on several state-of-the-art datasets and DoD benchmarks, the method generates fluent and accurate hard prompts that substantially outperform zero-shot baselines such as InSPEcT and even surpass few-shot learning performance in certain settings. This approach represents the first successful demonstration of high-quality, transferable conversion between soft and hard prompts, significantly enhancing both the interpretability and practical utility of prompt engineering.

Existing large language model (LLM) prompt datasets are fragmented across sources, heterogeneous in type, and lack systematic organization or analysis. Method: This paper introduces the first open-source, large-scale prompt dataset taxonomy covering diverse tasks, languages, and modalities; proposes a syntax-centered prompt optimization framework that leverages part-of-speech tagging and dependency parsing to generate syntactic embeddings, employs clustering to identify prototypical prompt representations, and guides LLMs to rewrite prompts toward these syntactic centers. Contribution/Results: The approach significantly improves prompt effectiveness and output semantic consistency. All datasets, preprocessing pipelines, and optimization code are publicly released, establishing foundational resources and methodological support for prompt engineering and human-AI interaction research.

Existing prompt engineering methods for input-sensitive tasks like machine translation predominantly optimize instructions rather than critical input components, resulting in limited generalizability and efficiency. Method: We propose the first lightweight, input-focused prompt rewriting framework: an end-to-end input rewriting network built upon a small-parameter language model (0.1B), trained via a novel back-translation–driven self-supervised objective to automatically refine source-language inputs. Crucially, no large language model is involved in training, significantly reducing computational overhead. Contribution/Results: Our approach achieves performance gains comparable to large-model–assisted prompt engineering—+2.3 BLEU on average—across multiple machine translation benchmarks. It introduces minimal parameter overhead while maintaining strong transferability to other input-dependent downstream tasks, offering an efficient, scalable alternative to instruction-centric prompting paradigms.