This article explains the fundamentals of prompt engineering for large language models, covering LLM output configuration, length and sampling controls, various prompt types, chain‑of‑thought and tree‑of‑thought reasoning methods, and practical best‑practice guidelines for creating high‑quality prompts.
The paper introduces T‑LLM, a time‑distillation framework that transfers predictive behavior from a lightweight teacher model to a general‑purpose LLM, enabling accurate multivariate time‑series forecasting across full‑sample, few‑shot, and zero‑shot settings while eliminating the need for large‑scale pre‑training.
This article chronicles the rapid progression of GPT models from the 2018 GPT‑1 pre‑training breakthrough through GPT‑2’s multitask learning, GPT‑3’s scaling laws and few‑shot abilities, to GPT‑4’s multimodal capabilities and the 2024 GPT‑4 Turbo, Sora, and GPT‑4o releases, while also explaining core LLM abilities and the decoder‑only architecture of GPT‑2.
This paper introduces ABPlanner, a few‑shot, context‑aware budget planner that enhances budget‑constrained auto‑bidding in online advertising by hierarchically allocating budgets across short‑term stages and training a sequential decision‑maker with deep reinforcement learning, achieving significant gains in simulated and real‑world A/B tests.
This article details the design and deployment of a multi‑layer LLM system that automatically reads massive creator group chats, extracts structured insights, mitigates hallucinations with dual‑model verification, uses few‑shot prompting for stable output, and delivers real‑time risk alerts and operational reports.
This paper presents a collaborative framework where a large language model generates high‑quality synthetic samples to augment a lightweight model, dramatically improving few‑shot user‑complaint intent recognition in 5G networks, achieving a 21% boost for rare categories and a 9% overall accuracy gain.
The paper introduces the LIMI framework, which achieves state‑of‑the‑art agent performance on AgencyBench using only 78 carefully crafted samples—outperforming baseline models trained on thousands of examples—by focusing on high‑quality, strategic data construction and demonstrating superior generalization across code, research, and tool‑use tasks.
This article reviews the MLLM4TS framework, which fuses visual representations of multivariate time series with large language models to address complex temporal dependencies, cross‑channel interactions, and task generalization, and demonstrates superior performance on classification, anomaly detection, forecasting, and few‑shot scenarios across multiple benchmarks.
Learn how to craft effective prompts for large language models by using a systematic structure—role and task, core principles, context handling, chain‑of‑thought, output specifications, and few‑shot examples—and discover techniques for generating and iteratively refining prompts with the model itself.
At SIGMOD 2025 in Berlin, Alibaba Cloud presented its paper “OpenSearch‑SQL: Enhancing Text‑to‑SQL with Dynamic Few‑shot and Consistency Alignment,” which introduces a self‑taught few‑shot mechanism and multi‑agent alignment to significantly improve NL2SQL query accuracy, alongside a keynote on AI search trends.
This article reviews recent AI self‑evolution research and provides an in‑depth analysis of the SEAL (Self‑Adapting Language) framework, which enables large language models to generate and learn from their own synthetic data through a nested reinforcement‑learning and fine‑tuning loop, with experimental results on few‑shot and knowledge‑integration tasks.
Distribution‑aware Graph Prompt Tuning (DAGPrompT) tackles the pre‑training/downstream mismatch on heterophilic graphs by jointly applying low‑rank GLoRA adaptation and hop‑specific prompts that recast tasks as link‑prediction, yielding up to 4.79% accuracy gains and an average 2.43% improvement in few‑shot node classification.
Visual‑RFT introduces a reinforcement‑fine‑tuning paradigm for large multimodal models, extending rule‑based reward strategies from text‑only LLMs to visual‑language tasks such as detection and grounding, and demonstrates strong few‑shot performance gains over traditional supervised fine‑tuning across multiple benchmarks.
This article presents a systematic framework for crafting effective prompts, detailing the universal prompt template, role definition, task decomposition, RAG integration, few‑shot examples, memory handling, and parameter tuning to enhance large language model performance across diverse applications.
This article introduces a novel few‑shot learning approach tailored for multi‑new‑class scenarios, discusses its background, problem definition, proposed parallel training framework, hierarchical fine‑tuning method, and presents extensive experiments demonstrating superior performance and computational efficiency.
FocusICL mitigates the reverse‑scaling of in‑context learning by masking irrelevant tokens and applying hierarchical batch attention, cutting attention complexity, and delivering consistent query focus that yields average accuracy gains of about 5 % across multiple LLMs and benchmarks.
This article explains what Prompt Engineering is, traces its development from early command‑based interactions to modern adaptive and multimodal prompting, details various prompting techniques such as zero‑shot, few‑shot, Chain‑of‑Thought, hallucination‑reduction methods, and demonstrates their practical use in a JD Logistics SKU piece‑type classification case with code examples.
This article presents a systematic, four‑part Prompt engineering framework—role definition, problem description, goal setting, and requirement specification—augmented with RAG, few‑shot examples, memory handling, and model‑parameter tuning, enabling developers to craft high‑quality prompts for large language models across diverse tasks.
The article presents a universal four‑part prompt template—role, problem description, goal, and requirements—augmented with role definitions, RAG‑based knowledge retrieval, few‑shot examples, memory handling, temperature/top‑p tuning, and automated optimization techniques such as APE, APO, and OPRO, enabling developers to reliably craft high‑quality prompts for LLMs.
This paper introduces M2SD, a dual‑branch multiple‑mixing self‑distillation framework that expands feature space, mitigates overfitting and catastrophic forgetting, and achieves state‑of‑the‑art results on CIFAR‑100, CUB‑200 and miniImageNet for few‑shot class‑incremental learning.
Alibaba Cloud’s AI platform PAI showcased four cutting‑edge papers at AAAI‑2024—introducing AMFormer for deep tabular learning via arithmetic feature interaction, MuLTI for efficient video‑language understanding, M2SD for few‑shot class‑incremental learning, and M2Doc for multi‑modal document layout analysis—demonstrating the platform’s growing impact on artificial‑intelligence research.
The article examines the security challenges introduced by large‑model AIGC, outlines three technical upgrade paths—richer training data, few‑shot model fine‑tuning, and multimodal fusion—and demonstrates practical implementations that dramatically improve detection efficiency, accuracy, and scalability.
This article reviews Alibaba DAMO Academy's research on information extraction, covering background concepts, PLM-era extraction paradigms, few‑shot extraction techniques, and the emerging LLM‑era approaches, while also sharing practical insights, benchmark results, and future directions.
This article introduces the AI sales assistant from 58.com, detailing its background, a few‑shot voice cloning approach using real dialogue data, multi‑accent naturalness optimization, deployment architecture, and future plans, while evaluating performance metrics and discussing challenges in speech synthesis quality and stability.
Prompt engineering teaches how to craft effective instructions, context, input data, and output formats for large language models, using clear commands, iterative refinement, and advanced methods such as zero‑shot, few‑shot, chain‑of‑thought, Tree of Thoughts, retrieval‑augmented and progressive‑hint prompting to achieve precise, reliable results across diverse tasks.
This article introduces the tasks, techniques, and frameworks for financial event analysis using pre‑trained language models, covering unstructured data parsing, event semantics, graph construction, detection, extraction, and prediction, and presents the TDE‑GTEE model that achieves state‑of‑the‑art performance even in few‑shot scenarios.
This article explains why effective prompt design—covering system messages, few‑shot learning, non‑dialogue scenarios, explicit instructions, output shaping, syntax cues, task decomposition, chain‑of‑thought, and real‑world context—is essential for reliable large language model results and provides practical examples and tips.
PaddleDetection v2.6 expands the PP‑YOLOE family with rotating, small‑object, dense‑object, and ultra‑lightweight edge‑GPU models, upgrades PP‑Human and PP‑Vehicle toolboxes, releases semi‑supervised, few‑shot and distillation learning methods, adds numerous state‑of‑the‑art algorithms, and improves infrastructure with Python 3.10, EMA filtering and AdamW support.
This article explains the concept of in‑context learning in large language models, presents experimental evaluations such as copy‑output, date‑formatting, and label‑remapping tasks, and discusses a recent theoretical analysis that links attention layers to implicit gradient‑based fine‑tuning, highlighting why model scale and data volume matter.
Unified Prompt Tuning (UPT) is a meta-learning based few‑shot algorithm that converts diverse NLP tasks into a common Prompt‑Options‑Verbalizer format, enabling large pre‑trained language models to achieve higher accuracy with minimal labeled data, as demonstrated on EMNLP‑2022 benchmarks and SuperGLUE datasets.
FSL++—a RoBERTa‑large‑based few‑shot model enhanced with domain‑adaptive pre‑training, prompt learning, diverse embedding‑level augmentations, and ensemble self‑training—topped the Chinese FewCLUE benchmark, beating human accuracy on news and scientific classification tasks and delivering measurable gains across multiple Meituan product scenarios.
This article introduces PaddleNLP's UIE tool for open‑domain information extraction, explains its underlying UIE model and ERNIE 3.0 foundation, showcases concise Python code for entity and event extraction, and presents few‑shot and SOTA performance results across multiple IE benchmarks.
This article explains why cold‑start is a critical challenge for recommender systems, outlines four practical strategies—generalization, fast data collection, transfer learning, and few‑shot learning—and then details the DropoutNet model, its end‑to‑end training, loss functions, negative‑sampling techniques, and open‑source implementation.
JD Research Institute and the University of Sydney introduced ViTAEv2, a 600‑million‑parameter deep learning model that achieved a world‑leading 91.2% top‑1 accuracy on ImageNet Real without external data, demonstrating strong few‑shot learning, reducing labeling costs, and promising advances across many computer‑vision tasks.
PaddleNLP v2.1 introduces an industrial‑grade Taskflow for eight NLP scenarios, a three‑line few‑shot learning paradigm that boosts small‑sample performance, and a FasterTransformer‑based inference engine that delivers up to 28‑fold speedup for text generation, all backed by extensive model and algorithm integrations.
Meituan’s NLP team leverages few‑shot learning—using data‑augmentation, semi‑supervised, ensemble/self‑training, and domain‑adaptation techniques—to cut annotation costs, achieving 1–2 percentage‑point accuracy gains on internal benchmarks and deploying high‑performing models for tasks such as topic classification, fake‑review detection, and sentiment analysis, while planning broader platform and model extensions.
This article presents Didi's exploration of few‑shot learning, data‑augmentation, semi‑supervised self‑training and multi‑task learning techniques to address the scarcity of labeled samples in safety and governance scenarios, demonstrating practical solutions and performance gains across various risk‑detection tasks.
To overcome scarce labeled data for safety and governance, Didi combines few‑shot learning with systematic data augmentation, self‑training semi‑supervised labeling, and multi‑task neural architectures, cutting labeling costs and reducing log‑loss by over 20% while boosting ROC‑AUC and PR‑AUC across harassment detection, expense‑complaint, and route‑intercept use cases.
Alibaba Entertainment’s AI‑driven workflow combines perception, cognition, and decision layers—leveraging knowledge graphs, few‑shot learning, and real‑time sentiment analysis across casting, production, and distribution—to predict content demand with sparse data, boosting prediction accuracy by 10%, quality by 20% and cutting low‑traffic titles.
This article reviews Alibaba DAMO’s Conversational AI team research on low‑resource few‑shot learning, introducing induction and dynamic‑memory networks, detailing their architecture, experimental setup on ARSC, ODIC and miniRCV1 datasets, and demonstrating state‑of‑the‑art performance improvements.
This article presents a meta‑learning based end‑to‑end task‑oriented dialogue system that quickly adapts to new scenarios with limited data and improves robustness through a human‑machine collaboration decision module, validated on extended‑bAbI benchmarks and real‑world Alibaba Cloud customer‑service applications.
This comprehensive guide examines real‑world text classification projects, covering label taxonomy design, data scarcity solutions, efficient annotation, new‑class discovery, algorithm selection, evaluation metrics, OOV handling, model evolution, rule‑model integration, performance‑boosting tricks, and inference under resource constraints.
The JD ARVR technology department announced two peer‑reviewed papers—one on a novel defocus blur detection network published in Transaction on Multimedia and another on a transductive relation‑propagation network for few‑shot learning accepted at IJCAI 2020—highlighting their advanced AI research and future AR‑VR ecosystem plans.
This article describes how advanced natural‑language‑processing, big‑data, and deep‑learning techniques are integrated into an end‑to‑end platform for financial asset management, covering large‑scale bid‑tender text analysis, few‑shot sentiment monitoring, model architectures, data‑enhancement methods, and practical deployment results.
Alibaba researcher Lei Yin explains how FashionAI transforms fashion e‑commerce by reconstructing industry knowledge, leveraging recommendation systems, knowledge graphs, and few‑shot learning to dramatically accelerate model training and improve attribute recognition accuracy.
This article reviews the evolution, architecture, and key technical challenges of general‑domain conversational QA systems, describing Alibaba UC’s search background, dialogue bot types, data pipelines, and advanced methods such as transfer learning, few‑shot learning, and multi‑dimensional dialogue management.
This article introduces meta‑learning (learning to learn), its historical roots, explains why it excels in small‑sample and multi‑task settings, contrasts it with supervised and reinforcement learning, and outlines the theoretical reasons it enables rapid few‑shot adaptation.
