Prompting Strategies for Large Language Models: A Structured Review of Techniques, Applications, and Open Challenges
ID：39
View protection：Participant Only
Updated time：2026-07-22 16:09:19 Views：25
Online
Abstract
Prompting strategies for large language models, including chain-of-thought reasoning, few-shot demonstrations, instruction tuning, and automatic prompt optimisation, have become central to how practitioners extract useful behaviour from these systems without modifying model weights. Recent growth in research on these methods has produced a wide and fragmented body of work that spans reasoning benchmarks, medical applications, code generation, and multilingual settings. This paper presents a comprehensive review of sixty-three studies published between 2021 and 2025, covering commonly examined technique categories, datasets, evaluation approaches, and reported limitations. Stepwise reasoning formats, in-context learning, zero-shot instruction following, soft prompt tuning, and agent-style iterative prompting have each shown meaningful improvements over baseline approaches across diverse task types. A systematic review process is used to compare performance trends and surface recurring limitations such as output sensitivity to phrasing, the absence of standardised benchmarks, prompt injection vulnerabilities, and limited multilingual coverage. The study draws attention to future directions including security-aware prompting, cross-lingual adaptation, explainable prompt analysis, and unified evaluation infrastructure. Taken together, the reviewed work demonstrates that prompt engineering has matured into a field with genuine technical depth, though several foundational problems remain unresolved.
Keywords
Prompt Engineering, Large Language Models, Chain-of-Thought Prompting, Few-Shot Learning, Instruction Tuning, Automatic Prompt Optimization, Agentic Prompting, Natural Language Processing
Post comments