Researchers from seven universities open-source OpenWAM, a modular world-action model stack combining video generation priors with robot action learning, featuring controlled experiments on architecture, information flow, and cross-embodiment pretraining, achieving state-of-the-art results on eight simulation benchmarks and real-robot tasks.
ToolCPT demonstrates that embedding real‑world tool knowledge during LLM pre‑training, rather than fine‑tuning, dramatically improves agent performance, using a mined corpus of 5.1 million proxy tools, detailed playbooks, and a 10 % tool‑data mix that yields up to 7.15‑point gains on benchmark tasks.
The article analyzes recent leaks about OpenAI’s upcoming large‑scale pre‑training model named Doug, situates it within the company’s post‑GPT‑4o scaling strategy that now relies on reinforcement learning and inference‑time compute, and assesses the competitive pressure from Google’s Gemini 3 and the implications of a potential base‑model overhaul.
The article explains that large‑model pre‑training is fundamentally a next‑word prediction task that forces the model to compress massive text corpora, discover statistical regularities, build semantic representations, and, at sufficient scale, exhibit emergent abilities, with practical implications for model selection and AI system design.
LingBot-VA 2.0 presents an industry‑first embodied‑native pre‑training model that aligns video prediction with action generation, uses a semantic‑visual‑action tokenizer, multi‑chunk prediction, foresight reasoning, and a sparse MoE architecture to achieve higher success rates and up to 6.5× faster end‑to‑end inference on real‑world robot tasks.
The article breaks down the exact five‑stage pipeline—data collection, pre‑training, supervised fine‑tuning, reward modeling, and reinforcement learning—that transforms raw internet text into powerful LLMs like GPT and Claude, and explains how understanding each step lets you build a miniature version yourself.
Keye‑VL‑2.0, an open‑source MoE multimodal foundation model, tackles hour‑level video understanding and agentic intelligence by embedding DeepSeek Sparse Attention into a GQA‑based architecture, enabling near‑lossless 256 K token context, four‑stage pre‑training, diverse RL distillation techniques, and achieving state‑of‑the‑art results on long‑video benchmarks, with weights publicly released.
OneReason introduces a systematic reasoning capability into industrial recommendation models through multi‑stage pre‑training, chain‑of‑thought fine‑tuning, and reinforcement learning, achieving significant gains in click‑through, revenue, and cross‑domain recommendation performance while preserving the underlying language abilities of the base model.
The OneReason report details how Kuaishou’s recommendation team injects reasoning into large‑scale recommender models through a four‑level pre‑training pipeline, chain‑of‑thought (CoT) fine‑tuning, and specialized reinforcement learning, achieving significant offline gains and a 10.33% exposure lift in a live A/B test.
The newly open‑sourced Wall‑OSS‑0.5 VLA model demonstrates that a large‑scale pre‑trained embodied robot brain can achieve strong zero‑shot performance on 17 real‑world tasks, exhibit staircase emergence with longer pre‑training, and far surpass the industry baseline after fine‑tuning, while also revealing current precision limits.
The article explains that large language models are massive probability engines built on the Transformer architecture with self‑attention, trained through costly pre‑training on trillions of tokens, then refined by instruction fine‑tuning and RLHF, ultimately predicting the next token to generate text.
Andrej Karpathy announced his return to frontline AI research by joining Anthropic just as both companies prepare for IPOs, a move that leverages his extensive background, reflects shifting LLM scaling priorities, and signals a strategic talent and technology win for Anthropic in the competitive AI market.
The article distinguishes two meanings of “end‑to‑end,” then outlines four sequential stages—defining data and scenario, massive pre‑training on audio‑text pairs, task alignment via instruction or supervised fine‑tuning, and optional preference tuning—to guide engineers in building usable speech assistants.
The article explains that pretraining for speech models functions like a broad liberal‑arts education, teaching universal acoustic and linguistic patterns through next‑token prediction, joint audio‑text training, and mask‑or contrast objectives, while clarifying common misconceptions and highlighting data bias and the need for clean, task‑specific fine‑tuning.
This article breaks down the fundamentals of large model training—covering data, parameters, neural networks, loss functions, gradient descent, pre‑training, and fine‑tuning—in plain language so readers can grasp how massive models learn without needing to dive into complex mathematics.
The article explains how pretraining on massive generic data creates a reusable base model, while fine‑tuning uses smaller, high‑quality task‑specific data to adapt the model, covering objectives, data scale, cost, methods, and why most projects prefer fine‑tuning.
The article examines how the dramatic increase in GPU compute power—illustrated by a single H100 GPU equaling about 200 Hadoop instances—challenges the dominance of tree‑based models for structured data, presents scaling‑law experiments with KMLP and FOUND, and argues that pre‑training can redefine the balance between compute, data, and algorithms.
The article examines how a single NVIDIA H100 GPU delivers roughly 200‑fold more FP16 compute than a 96‑core CPU Hadoop node, explores the "Bitter Lesson" of scaling‑driven AI breakthroughs, and presents large‑scale pretraining experiments that show table and sequence models now exhibit clear scaling laws, challenging the dominance of traditional tree‑based approaches.
This article explains the full lifecycle of large language models in 2026, covering pretraining fundamentals, the limits of classic Scaling Laws, data‑centric advances, fine‑tuning strategies, RLHF, DPO, and the emerging post‑training methods GRPO, DAPO and RLVR, with concrete benchmarks and cost analyses.
This guide breaks down the four major large‑model training paradigms—pre‑training, supervised fine‑tuning, preference alignment, and RLHF—explaining which parameters are updated, how attention is reshaped, and what capabilities are gained, so you can deliver a structured, interview‑ready answer.
A 630‑line Python Autoresearch project sparked a community‑run distributed system that created over 80 autonomous AI agents, executed more than 2,300 experiments in four days, self‑organized roles and peer‑review, and uncovered ten concrete pre‑training findings.
Zitong Yang’s Stanford PhD defense introduced “continually self‑improving AI,” a system that autonomously refines its own parameters, generates synthetic training data, and even designs its own learning algorithms, with experiments on synthetic continual training, synthetic‑bootstrap pre‑training, and AI‑design‑AI demonstrating measurable gains over static baselines.
The article reviews the HORAI model, which introduces a frequency‑enhanced multimodal pretraining paradigm and the massive MM‑TS dataset, showing that integrating derived images, endogenous text, and real‑world news dramatically improves zero‑shot forecasting and anomaly detection across six domains.
This article breaks down the three core stages of large language model training—pretraining, supervised fine‑tuning, and alignment—detailing their objectives, typical data formats, scale requirements, and the latest techniques such as RLHF and DPO.
Recent research shows that applying reinforcement learning to large language models can dramatically improve inference performance, but its effectiveness depends on the token distribution produced during pre‑training, prompting a novel rewrite of cross‑entropy as a single‑step policy gradient with controllable entropy parameters.
This extensive guide walks readers through the fundamentals of large language models, covering transformer architecture, pre‑training and fine‑tuning techniques, scaling laws, emergent abilities, mixture‑of‑experts designs, and practical comparisons, providing clear explanations, code snippets, and visual illustrations for deep learning practitioners.
The article demystifies large language models for frontend developers by likening token prediction to autocomplete, explaining tokens, context windows, temperature, the two-stage training process, and the critical role of prompts, using concrete code examples and analogies to familiar frontend concepts.
The X‑VLA paper introduces a 0.9‑billion‑parameter, fully open‑source embodied model that uses a learnable soft‑prompt and divide‑and‑conquer encoding to handle heterogeneous robot vision inputs, achieving a record‑breaking 120‑minute autonomous clothing‑folding task while surpassing benchmarks across five simulation environments.
This weekly roundup highlights five recent AI research papers—including CoCa’s contrastive captioning model, the Game‑TARS framework for scalable game agents, Kimi Linear’s efficient attention architecture, the Continuous Autoregressive Language Model (CALM), and a comprehensive survey of Context Engineering—summarizing their core contributions and providing direct links.
TripCast introduces a masked 2D transformer pre‑training framework that treats travel demand as a two‑dimensional time‑series problem, leveraging time‑patch tokenization, dual masking and RevIN normalization to achieve state‑of‑the‑art forecasting performance on massive real‑world travel data.
This paper introduces IAK, a unified multi‑domain recommendation paradigm that treats the system as a large model, leveraging pre‑training and fine‑tuning with an information‑aware adaptive kernel to capture rapid user interest shifts while reducing training costs and improving online performance.
This guide walks through the complete end‑to‑end process of training a large language model from scratch, covering data collection, cleaning, tokenization, pre‑training objectives and engineering, post‑training alignment methods, scaling laws, over‑fitting mitigation, and gradient‑stability techniques.
Keye‑VL‑1.5, an 8‑billion‑parameter multimodal large language model, introduces a Slow‑Fast video encoding strategy, a four‑stage progressive pre‑training pipeline with 128K context, and a sophisticated post‑training regime that together achieve state‑of‑the‑art performance on video and vision‑language benchmarks while maintaining strong general capabilities.
This article summarizes recent AI research on quantitative finance, covering AlphaAgent's LLM-driven alpha mining, UMI's multi‑level irrationality factors, PDU's progressive dependency learning for stock ranking, SSPT's stock‑specific pretraining transformer, and Enhancer's distribution‑aware meta‑learning framework, all of which demonstrate improved stock prediction and resistance to alpha decay.
This article reviews the SSPT paper, which introduces three stock‑specific pre‑training tasks—stock code classification, sector classification, and moving‑average prediction—built on a two‑layer Transformer, and demonstrates through extensive experiments across five market datasets that these tasks consistently improve cumulative return and Sharpe ratio over baselines.
This article surveys recent large‑scale corpus rewriting techniques for LLM pre‑training, covering K2’s token‑utilization strategies, domain‑specific methods like SwallowMath/Code, reStructured pretraining, the WRAP pipeline, Nemotron‑CC filtering, Pro‑X noise removal, and the MAGA multi‑style expansion, while highlighting challenges, experimental findings, and open research questions.
The USP framework introduces masked latent modeling within a VAE space to pre‑train ViT encoders, enabling seamless weight transfer to both image classification, segmentation, and diffusion‑based generation tasks, dramatically speeding up DiT and SiT models while preserving strong visual representations.
The USP framework introduces masked latent modeling within a VAE space to pretrain ViT encoders, enabling seamless weight transfer to both image classification and diffusion‑based generation tasks, dramatically accelerating training while preserving strong performance across multiple benchmarks.
Pre‑training enables AI models to first acquire a universal knowledge map from massive unlabelled text, then quickly adapt to specific tasks with minimal labelled data, offering superior generalization, reduced annotation costs, and versatile applications across chatbots, content creation, retrieval, coding assistance, and more.
AI applicationsPretrainingSelf-Supervised Learning
The article details Qwen3's three‑phase pretraining pipeline, long‑context extensions, a cold‑start long‑chain‑of‑thought dataset, reinforcement‑learning fine‑tuning with custom rewards, and a two‑stage distillation process that yields versatile, thought‑controlled language models.
This article provides a detailed Chinese‑to‑English summary of Andrej Karpathy’s 7‑hour LLM tutorial, covering chat process analysis, tokenization, pre‑training data pipelines, model architecture, training strategies, post‑training fine‑tuning, reinforcement learning, chain‑of‑thought reasoning, and current industry applications.
SimpleAR demonstrates that a vanilla autoregressive model with only 0.5 B parameters can generate high‑fidelity 1024×1024 images, covering pretraining, supervised fine‑tuning, and reinforcement learning, achieving competitive GenEval (0.59) and DPG‑Bench (79.66) scores while reducing inference time to about 14 seconds with vLLM and KV‑cache optimizations.
SimpleAR is a minimalist autoregressive visual generation framework that, with only 0.5 B parameters, achieves competitive 1024×1024 image synthesis through a three‑stage pipeline of large‑scale pretraining, supervised fine‑tuning, and GRPO‑based reinforcement learning, and demonstrates significant inference speedups using KV‑cache, vLLM, and speculative decoding.
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.
The article outlines the three development phases of large language models—building, pre‑training, and fine‑tuning—describes usage options, highlights key factors such as data scale, architecture, training processes, and evaluation, and offers practical advice for cost‑effective development.
GLM introduces a unified pretraining framework that combines autoregressive blank‑filling with 2D positional encoding and span‑shuffle, achieving superior performance over BERT, T5 and GPT on a range of NLU and generation tasks such as SuperGLUE, text‑filling, and language modeling.
GLM introduces a universal language model that combines autoregressive blank‑filling with 2D positional encoding and span‑shuffle training, achieving superior performance over BERT, T5, and GPT across NLU, conditional and unconditional generation tasks, as demonstrated on SuperGLUE and other benchmarks.
The article explains DeepSeek's two‑stage training pipeline—including massive pre‑training on trillions of tokens and post‑training via instruction tuning and reinforcement learning from human feedback—describes the differences between its V3 instruction model and R1 reasoning model, and highlights performance optimizations and emerging research directions.
This presentation details engineering‑focused approaches to building, mixing, and filtering data for large language models, covering data preparation, pre‑training mix strategies such as DoReMi, DoGE and online sampling, post‑training data quality selection methods, and practical Q&A on scaling laws and PDF processing.
This article examines the unique multimodal and multi‑granular nature of web pages, compares fusion strategies, proposes a cross‑modal attention approach, outlines fine‑ and coarse‑grained pre‑training tasks, and explores low‑cost adaptor methods for adapting large multimodal models to web‑page modeling in the LLM era.
The Qwen2.5 Technical Report introduces a new series of large language models with up to 72 B parameters, expanded pre‑training data to 18 trillion tokens, advanced supervised fine‑tuning and reinforcement learning pipelines, and demonstrates strong performance across comprehension, reasoning, coding, and long‑context tasks.
This issue reviews the latest AI breakthroughs—including OpenAI’s O3 and o1 models, pricing cuts, new features in ChatGPT, product launches like Pika 2.0 and Gemini 2.0, a heated debate on pre‑training data bottlenecks sparked by Ilya Sutskever, a novel black‑box forgetting method, and DeepMind’s Genie 2 3D world generator—highlighting how industry dynamics and research directions are reshaping the field.
This article explains fine‑tuning in machine learning, covering its definition, why it matters, the role of pre‑trained models, detailed step‑by‑step procedures, advantages, and diverse applications such as NLP, computer vision, speech and finance, with practical examples like face recognition and object detection.
This article details the end‑to‑end process of pre‑training, fine‑tuning, and evaluating a 1‑billion‑parameter Chinese LLM named Steel‑LLM on limited hardware, covering data collection, pipeline design, training framework choices, architectural tweaks, performance results, and practical lessons for resource‑constrained developers.
Bilibili’s in‑house role‑playing large language model, built on the Index architecture and refined through pre‑training, supervised fine‑tuning, and preference optimization (PPO and DPO), achieved top scores on the Chinese CharacterEval benchmark, surpassing rivals while incorporating safety alignment and showcasing consistent, personality‑driven dialogue examples.
The article traces the development of GPT series—from the 2017 Transformer breakthrough, through GPT‑1, GPT‑2, and GPT‑3’s 175 billion parameters, to later models like Codex and ChatGPT—highlighting key papers, architectural choices, and the surprising role of OpenAI’s decoder‑only approach.
An extensive review of Meta’s Llama 3 model breaks down its pre‑training data pipeline, scaling laws, architectural tweaks like GQA and RoPE, post‑training methods such as SFT, DPO, and reward modeling, and evaluates benchmark results, offering practical insights for researchers and engineers building large language models.
Index-1.9B-32K is a 1.9B-parameter model with a 32K token context window, achieving strong long‑text performance comparable to larger models while using only about 2% of GPT‑4’s compute, trained via long pre‑training and supervised fine‑tuning, with a trade‑off of reduced short‑context ability.
The answer explains that joining a pre‑training team accelerates a newcomer’s engineering abilities through hands‑on work with large‑scale data pipelines, distributed training code, and debugging, while SFT teams focus mainly on data labeling, making pre‑training the more effective path for rapid skill growth.
This article surveys data management for large language model training, covering an overview, pre‑training data composition, scaling‑law‑driven quantity control, quality filtering, deduplication, harmful‑content removal, instruction fine‑tuning strategies, dynamic data selection, and emerging research challenges such as bias mitigation, multimodal data handling, and synthetic‑data filtering.
The answer argues that fresh graduates should join pre‑training teams because the required engineering tasks—large‑scale data crawling, Hadoop/Spark pipelines, torch and CUDA setup, megatron code debugging, and scaling‑law experiments—rapidly sharpen coding skills, while SFT work focuses mainly on data labeling and offers slower technical growth.
This article presents a comprehensive study on integrating multimodal image‑text representations into large‑scale e‑commerce advertising CTR models, introducing a semantic‑aware contrastive pre‑training (SCL) method and two application algorithms (SimTier and MAKE) that together achieve over 1 % GAUC improvement and significant online gains.
The article provides an in‑depth analysis of LLaMA 3 405B, covering its dense Transformer architecture, three‑stage pre‑training (initial, long‑context, annealing), iterative post‑training with RM‑guided rejection sampling, the decision against MOE, and the broader implications for both large and small model development.
This article provides a comprehensive technical overview of the ChatGLM series—from the original ChatGLM‑6B model and its GLM‑based pre‑training framework to the enhancements in ChatGLM‑2, the architectural parity of ChatGLM‑3, and the advanced capabilities of the latest ChatGLM‑4, covering model structure, position encoding, attention mechanisms, multi‑task pretraining, and tool integration.
The report presents the open‑source Index‑1.9B family—base, pure, chat, and character variants—detailing benchmark results, pre‑training optimizations such as a normalized LM‑Head and deeper‑slim architectures, the importance of modest instruction data, alignment via SFT/DPO, role‑play enhancements with RAG, and acknowledges remaining safety and factual limitations.
This article analyzes whether large‑scale web crawls, when meticulously filtered and deduplicated, can match or surpass the performance of high‑quality curated datasets in training large language models, covering dataset composition, processing pipelines, experimental results, scaling‑law implications, and future data‑efficiency strategies.
This step‑by‑step guide shows how to use Hugging Face's Trainer API to pre‑train an ultra‑small LLaMA‑3 model (under 20 M parameters) on the TinyStories dataset, covering model configuration, tokenizer setup, data preprocessing, collators, training arguments, and inference results.
This article presents Ant Group's multimodal research on video retrieval, detailing video‑text semantic search and video‑video same‑source search, introducing a large Chinese pre‑training dataset, novel pre‑training, hard‑sample mining, fine‑grained modeling techniques, and an efficient end‑to‑end copyright detection framework.
This article presents Baidu's exploration of multimodal content understanding for commercial advertising, detailing the ViCAN pre‑training model, its contrastive and mask‑language learning tasks, integration across recall, ranking and risk‑control pipelines, quantization with MMDict, and future AIGC‑driven generation, all backed by extensive experiments and Q&A.
This article reviews Tongyi Lab's work on the OFA framework for generative multimodal pretraining and the ONE-PEACE model for unified multimodal representation learning, detailing their architectures, training strategies, experimental results across vision‑language and audio tasks, and future research directions.
This article provides a comprehensive technical overview of Meta's Llama 2 series, covering its architectural upgrades such as Group Query Attention, the pre‑training dataset and hyper‑parameters, loss behavior, benchmark comparisons, and the supervised fine‑tuning pipeline with safety considerations.
The paper presents MuLTI, a multimodal video‑language model that tackles the memory and efficiency challenges of long video‑text sequences by introducing a Text‑Guided MultiWay Sampler and a Multiple Choice Modeling pre‑training task, achieving state‑of‑the‑art results on video QA and retrieval while drastically reducing GPU memory consumption.
This article explains the concepts of pretraining and fine‑tuning for large language models, compares full‑parameter, LoRA and QLoRA approaches, discusses resource consumption, introduces the ModelScope SWIFT framework with code examples, and shows how fine‑tuning can improve data‑visualisation tasks while reducing token usage.
This article provides a comprehensive English overview of BERT, covering its original paper, model architecture, pre‑training objectives (Masked Language Model and Next Sentence Prediction), differences from ELMo, GPT and vanilla Transformers, parameter counts, main contributions, and a range of NLP application scenarios such as text classification, sentiment analysis, NER, and machine translation.
This article provides a comprehensive overview of Google's T5 model, detailing its unified text‑to‑text formulation, encoder‑decoder architecture, three model variants, attention mask designs, training strategies, model sizes, experimental results, and key contributions to natural language processing.
The paper proposes a multi‑scale stochastic distribution prediction (MSDP) framework that learns robust user behavior representations by predicting behavior distributions over random time windows, incorporates contrastive regularization, and demonstrates superior performance on both proprietary financial risk data and a public e‑commerce dataset compared with existing masked and next‑behavior pre‑training methods.
This article presents a comprehensive overview of how natural language processing is applied to financial risk control, covering text and non‑text sequence modeling, tokenization strategies, transformer‑based long‑text architectures, multi‑modal fusion methods, pre‑training techniques and practical sample‑optimization approaches.
This article reviews the evolution of the GPT family from GPT‑1 to GPT‑3, comparing encoder‑decoder architectures, explaining the shift from supervised fine‑tuning to zero‑shot and few‑shot learning, and highlighting the architectural and training innovations that enabled large‑scale language models.
This article provides a comprehensive overview of the attention mechanism, its mathematical foundations, the transformer model architecture—including encoder and decoder components—and the BERT pre‑training model, detailing their principles, implementations, and applications in natural language processing.
This article presents OPPO Research Institute's recent advances in large‑model AI, detailing the CHAOS pretrained model that topped the CLUE leaderboard, the knowledge‑enhanced training pipeline, and the GammaE model for multi‑hop reasoning over knowledge graphs, together with experimental results and practical training tips.
This article presents a low‑resource NLP pretraining approach that combines transformer‑based language modeling with contrastive vector learning, details the unsupervised sample‑pair construction, introduces a camel‑shaped masking distribution, and demonstrates through extensive experiments that the resulting model achieves strong zero‑shot NLU, NLG, and retrieval performance while requiring minimal compute and data.
The article surveys recent academic advances in both sparse inverted‑index and dense semantic retrieval for large‑scale search, highlighting key papers on decision frameworks, benchmarks, sparse lexical models, dual encoders, and hybrid techniques, while discussing challenges such as single‑vector limits and proposing multi‑view and hybrid solutions.
This article reviews recent AIGC progress, introducing the AliceMind series of text generation models—including PALM, PLUG, and a Chinese GPT‑3—alongside the multimodal mPLUG architecture, and discusses their training strategies, performance results, and practical deployment insights.
This article details Tencent's development of the 1‑trillion‑parameter HunYuan‑NLP model, covering its MoE architecture, cost‑effective pre‑training strategies, distributed training framework, model compression toolkit, and successful deployment across advertising, gaming, and other Tencent services.
To boost NetEase Cloud Music’s comment handling, the authors combine active‑learning‑driven relabeling, domain‑specific MLM pretraining, contrastive‑learning‑based sample expansion, and multi‑task BERT sharing to raise sentiment‑classification precision and recall above 90 % and double moderation clean‑rate, while employing prompt‑generated story themes, IP‑focused classifiers, and hot‑word aggregation for effective short‑text topic clustering and scalable, theme‑aware distribution.
This article presents a comprehensive overview of multimodal pre‑training, describing its motivation, architecture choices, large‑scale Chinese image‑text dataset construction, training optimizations, performance benchmarks, downstream applications, and a Q&A session that highlights practical deployment considerations.
Prof. Qiu Xipeng’s talk highlighted how large language models can be offered as a service and efficiently adapted via in‑context learning, lightweight label‑tuning, and gradient‑free black‑box optimization, showcasing a unified asymmetric Transformer (CPT) that handles understanding, generation, ABSA and NER tasks while reducing resource demands.
CKBERT, a Chinese knowledge‑enhanced BERT developed by Alibaba’s EasyNLP team, integrates external knowledge graphs and internal linguistic cues through novel pre‑training tasks, offers three model sizes compatible with HuggingFace and PAI, and demonstrates superior performance on CLUE and NER benchmarks while providing easy deployment on cloud platforms.
This article introduces the importance of image‑text cross‑modal representation, presents the Chinese Zero dataset with two pre‑training subsets and five downstream tasks, describes the R2D2 dual‑tower‑plus‑single‑tower pre‑training framework with multiple loss functions, and reports extensive experiments and real‑world deployment insights.
This article explains the importance of text generation, introduces the PEGASUS model’s gap‑sentence pre‑training for abstractive summarization, and shows how the EasyNLP framework integrates PEGASUS and other Chinese and English summarization models with step‑by‑step installation, data preparation, and training commands.
This talk presents recent year’s research on natural language generation, covering the ProphetNet pre‑trained generation model, external‑knowledge integration for generation, non‑autoregressive pre‑training (BANG), the Poolingformer long‑text architecture, EL‑attention for faster decoding, and a new multi‑task generation benchmark.
The OPPO XiaoBu team introduced OBERT, a series of 100M‑, 300M‑, and 1B‑parameter pretrained language models that leverage massive TB‑scale corpora, multi‑granular masking, retrieval‑augmented training, and distributed acceleration to achieve state‑of‑the‑art results on CLUE and KgCLUE benchmarks while enabling efficient industrial deployment.
This article reviews large‑scale image and multimodal pre‑training techniques—including contrastive learning, self‑supervised reconstruction, and multimodal alignment—explains data acquisition, model construction, evaluation metrics, and demonstrates how these methods are applied and optimized for real‑world information‑flow services.
The paper introduces MUVCOG, a heterogeneous graph neural network that models multi‑modal, multi‑query search sessions on Mobile Taobao by jointly learning attention‑based global and hierarchical local views through contrastive pre‑training, yielding universal session embeddings that markedly improve CTR prediction, query recommendation, and intent classification.
This article reviews recent advances in combining knowledge graphs with neural networks, covering generative pre‑training of graph neural networks, wiki‑graph based open‑domain question answering, differentiable logical reasoning, and a fuzzy‑logic query‑embedding model that improves performance on sparse‑relation queries.
The paper proposes a multimodal pre‑training framework that jointly encodes query text and item titles with images via shared and single‑stream towers, using MLM, MPM, QIC, and matching tasks, and demonstrates substantial Recall@K gains on a billion‑item e‑commerce catalog by leveraging visual cues to bridge the semantic gap.
This article reviews the motivation, methods, and results of graph-based self‑supervised pretraining for molecular data, introduces the ChemRL‑GEM model that incorporates 3‑D structural information, and demonstrates its superior performance on ADMET, affinity prediction, and benchmark competitions using the PaddleHelix platform.
This article traces the history of deep‑learning pre‑training techniques, comparing the parallel developments in natural‑language processing and computer vision—from early word2vec and bag‑of‑words models through ELMo and BERT to recent transformer‑based vision models like iGPT, ViT, BEiT and MAE—highlighting key innovations, challenges, and the convergence of the two fields.
This article traces the evolution of deep‑learning pre‑training techniques, starting with word2vec in NLP, moving through ELMo and BERT, then shifting to computer‑vision models such as iGPT, ViT, BEiT, and MAE, highlighting key innovations, challenges, and the convergence of NLP and CV paradigms.
This article reviews recent advances in user representation learning for recommender systems, covering self‑supervised pre‑training, lifelong learning, multi‑task modeling, and large‑scale contrastive methods, and provides code and dataset links for key papers such as PeterRec, Conure, DUPN, ShopperBERT, PTUM, UPRec, and LURM.