Long-context processing remains a significant challenge for large language models (LLMs). Retrieval-augmented generation (RAG) has recently emerged as a promising approach, enabling LLMs to selectively access relevant information from extended contexts to improve efficiency. However, existing RAG approaches often lag behind other efficient long-context processing methods primarily due to inherent limitations on inaccurate retrieval and fragmented contexts. To address these limitations, we propose RetroLM, a novel RAG framework designed for effective long-context processing. Unlike traditional approaches, RetroLM introduces KV-level retrieval augmentation, which partitions the LLM's KV cache into contiguous pages and performs encoding and decoding operations based on the retrieved KV pages. Built upon this framework, we further develop a specialized retriever for precise retrieval of critical pages and conduct unsupervised post-training to optimize the model’s ability to leverage retrieved information. Compared with traditional RAG, the new approach enhances robustness to retrieval inaccuracy, facilitates effective utilization of fragmented contexts, and saves the cost from repeated context-encoding operations. We conduct extensive evaluations across several popular benchmarks, including LongBench, InfiniteBench, and RULER. RetroLM consistently outperforms existing long-LLMs and RAG-based methods, especially in tasks requiring deep reasoning or extreme context lengths.

Kun Luo, Hongjin Qian, Zheng Liu, Ziyi Xia, Shitao Xiao, Zhao Cao, Siqi Bao, Jun Zhao, Kang Liu. Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2026.

Xingrun Xing, Zheng Liu, Shitao Xiao, Boyan Gao, Yiming Liang, Haokun Lin, Xianlin Zeng, Guoqi Li, Jiajun Zhang. Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2026.

Despite the rapidly growing demand for multimodal retrieval, progress in this field remains severely constrained by a lack of training data. In this paper, we introduce MegaPairs, a novel data synthesis method that leverages vision language models (VLMs) and open-domain images, together with a massive synthetic dataset generated from this method. Our empirical analysis shows that MegaPairs generates high-quality data, enabling the multimodal retriever to significantly outperform the baseline model trained on 70\times more data from existing datasets. Moreover, since MegaPairs solely relies on general image corpora and open-source VLMs, it can be easily scaled up, enabling continuous improvements in retrieval performance. In this stage, we produced more than 26 million training instances and trained several models of varying sizes using this data. These new models achieve state-of-the-art zero-shot performance across 4 popular composed image retrieval (CIR) benchmarks and the highest overall performance on the 36 datasets provided by MMEB. They also demonstrate notable performance improvements with additional downstream fine-tuning. Our code, synthesized dataset, and pre-trained models are publicly available at https://github.com/VectorSpaceLab/MegaPairs.

Jianlyu Chen, Nan Wang, Chaofan Li, Bo Wang, Shitao Xiao, Han Xiao, Hao Liao, Defu Lian, Zheng Liu. Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2025.

Zheng Liu, Ze Liu, Zhengyang Liang, Junjie Zhou, Shitao Xiao, Chao Gao, Chen Jason Zhang, Defu Lian. Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2025.

Recent advancements in text-to-image generation, notably the series of Stable Diffusion methods, have enabled the production of diverse, high-quality photo-realistic images. Nevertheless, these techniques still exhibit limitations in terms of knowledge access. Retrieval-augmented image generation is a straightforward way to tackle this problem. Current studies primarily utilize coarse-grained retrievers, employing initial prompts as search queries for knowledge retrieval. This approach, however, is ineffective in accessing valuable knowledge in long-tail text-to-image generation scenarios. To alleviate this problem, we introduce FineRAG, a fine-grained model that systematically breaks down the retrieval-augmented image generation task into four critical stages: query decomposition, candidate selection, retrieval-augmented diffusion, and self-reflection. Experimental results on both general and long-tailed benchmarks show that our proposed method significantly reduces the noise associated with retrieval-augmented image generation and performs better in complex, open-world scenarios.

The evaluation of Long Video Understanding (LVU) performance poses an important but challenging research problem. Despite previous efforts, the existing video understanding benchmarks are severely constrained by several issues, especially the insufficient lengths of videos, a lack of diversity in video types and evaluation tasks, and the inappropriateness for evaluating LVU performances. To address the above problems, we propose a new benchmark called MLVU (Multitask Long Video Understanding Benchmark) for the comprehensive and in-depth evaluation of LVU. MLVU presents the following critical values: 1) The substantial and flexible extension of video lengths, which enables the benchmark to evaluate LVU performance across a wide range of durations. 2) The inclusion of various video genres, such as movies, surveillance, egocentric videos, and cartoons, reflects the models’ LVU performances in different scenarios. 3) The development of diversified evaluation tasks, which enables a comprehensive examination of MLLMs’ key abilities in long-video understanding. The empirical study with 23 latest MLLMs reveals significant room for improvement in today’s technique, as all existing methods struggle with most of the evaluation tasks and exhibit severe performance degradation when handling longer videos. Additionally, it suggests that factors such as context length, image-understanding ability, and the choice of LLM backbone can play critical roles in future advancements. We anticipate that MLVU will advance the research of LVU by providing a comprehensive and in-depth analysis of MLLMs. The code and dataset can be accessed from https://github.com/JUNJIE99/MLVU.

The emergence of Large Language Models (LLMs) has unified language generation tasks and revolutionized human-machine interaction. However, in the realm of image generation, a unified model capable of handling various tasks within a single framework remains largely unexplored. In this work, we introduce OmniGen, a new diffusion model for unified image generation. OmniGen is characterized by the following features: 1) Unification: OmniGen not only demonstrates text-to-image generation capabilities but also inherently supports various downstream tasks, such as image editing, subject-driven generation, and visual-conditional generation. 2) Simplicity: The architecture of OmniGen is highly simplified, eliminating the need for additional plugins. Moreover, compared to existing diffusion models, it is more user-friendly and can complete complex tasks end-to-end through instructions without the need for extra intermediate steps, greatly simplifying the image generation workflow. 3) Knowledge Transfer: Benefit from learning in a unified format, OmniGen effectively transfers knowledge across different tasks, manages unseen tasks and domains, and exhibits novel capabilities. We also explore the model’s reasoning capabilities and potential applications of the chain-of-thought mechanism. This work represents the first attempt at a general-purpose image generation model, and we will release our resources at https://github.com/VectorSpaceLab/OmniGen to foster future advancements.

Large language models (LLMs) with decoder-only architectures have demonstrated exceptional text-generation capabilities across a variety of tasks. Some researchers have also adapted these models for text representation tasks. However, in text representation tasks, these models often face performance degradation on unseen tasks. In-context learning (ICL), which leverages examples provided in the input context, enables LLMs to handle unseen tasks effectively. Inspired by this, we aim to fully utilize the inherent properties of LLMs to enhance text representation performance across different tasks through the ICL approach. In this paper, we introduce a simple yet effective training strategy, which significantly improves text representation capabilities. Unlike previous models that prepend task instructions to the text, our method randomly samples a varying number of examples during training, endowing the embedding model with in-context learning abilities while maintaining its zero-shot capabilities. This approach does not require additional data construction or modifications to the model architecture. On the contrary, we find that some popular modifications to the model, such as bidirectional attention, can degrade performance, undermining the inherent characteristics of LLMs. We have publicly released our method at this \href{https://github.com/FlagOpen/FlagEmbedding}{repo}.

Recent advancements in large language models (LLMs) with billions of parameters have improved performance in various applications, but their inference processes demand significant energy and computational resources. In contrast, the human brain, with approximately 86 billion neurons, is much more energy-efficient than LLMs with similar parameters. Inspired by this, we redesign 7$\sim$70 billion parameter LLMs using bio-plausible spiking mechanisms, emulating the efficient behavior of the human brain. We propose the first spiking large language model, SpikeLLM. Coupled with the proposed model, two essential approaches are proposed to improve spike training efficiency: Generalized Integrate-and-Fire (GIF) neurons to compress spike length from $T$ to $\frac{T}{L} \log_2 L$ bits, and an Optimal Brain Spiking framework to divide outlier channels and allocate different $T$ for GIF neurons, which further compresses spike length to approximate $log_2T$ bits. The necessity of spike-driven LLM is proved by comparison with quantized LLMs with similar operations. In the OmniQuant pipeline, SpikeLLM reduces 11.01\% WikiText2 perplexity and improves 2.55\% accuracy of common scene reasoning on a LLAMA-7B W4A4 model. In the GPTQ pipeline, SpikeLLM achieves direct additive in linear layers, significantly exceeding PB-LLMs. Our code is publicly available at https://github.com/Xingrun-Xing2/SpikeLLM.

Long context compression is a critical research problem due to its significance in reducing the high computational and memory costs associated with LLMs. In this paper, we propose Activation Beacon, a plug-in module for transformer-based LLMs that targets effective, efficient, and flexible compression of long contexts. To achieve this, our method introduces the following technical designs. 1) We directly compress the activations (i.e. keys and values at every layer), rather than leveraging soft prompts to relay information (which constitute a major bottleneck to encapsulate the complex information within long contexts). 2) We tailor the compression workflow, where each fine-grained input unit is progressively compressed, enabling high-quality compression and efficient computation during both training and inference. 3) We train the model through compression-based auto-regression, making full use of plain texts and instructional data to optimize the model's compression performance. 4) During training, we randomly sample a compression ratio at each step, teaching the model to support a wide range of compression configurations. Extensive evaluations are conducted on various long-context tasks whose lengths (e.g., 128K) may far exceed the maximum training length (20K), such as document understanding, few-shot learning, and Needle-in-a-Haystack. Whilst existing methods struggle to handle these challenging tasks, Activation Beacon maintains a comparable performance to the uncompressed baseline across various scenarios, achieving a 2x acceleration in inference time and an 8x reduction of memory costs for KV cache.

The existing Retrieval-Augmented Generation (RAG) systems face significant challenges in terms of cost and effectiveness. On one hand, they need to encode the lengthy retrieved contexts before responding to the input tasks, which imposes substantial computational overhead. On the other hand, directly using generic Large Language Models (LLMs) often leads to sub-optimal answers, while task-specific fine-tuning may compromise the LLMs' general capabilities. To address these challenges, we introduce a novel approach called FlexRAG (Flexible Context Adaptation for RAG). In this approach, the retrieved contexts are compressed into compact embeddings before being encoded by the LLMs. Simultaneously, these compressed embeddings are optimized to enhance downstream RAG performance. A key feature of FlexRAG is its flexibility, which enables effective support for diverse compression ratios and selective preservation of important contexts. With these designs, FlexRAG achieves superior generation quality while significantly reducing running costs. The experiments across multiple QA datasets validate our approach as a cost-effective and flexible solution for RAG systems (codebase: https://github.com/wcyno23/FlexRAG).

Large language models (LLMs) provide powerful foundations to perform fine-grained text re-ranking. However, they are often prohibitive in reality due to constraints on computation bandwidth. In this work, we propose a flexible architecture called Matroyshka Re-Ranker, which is designed to facilitate runtime customization of model layers and sequence lengths at each layer based on users' configurations. Consequently, the LLM-based re-rankers can be made applicable across various real-world situations. The increased flexibility may come at the cost of precision loss. To address this problem, we introduce a suite of techniques to optimize the performance. First, we propose cascaded self-distillation, where each sub-architecture learns to preserve a precise re-ranking performance from its super components, whose predictions can be exploited as smooth and informative teacher signals. Second, we design a factorized compensation mechanism, where two collaborative LoRA modules, vertical and horizontal, are jointly employed to compensate for the precision loss resulted from arbitrary combinations of layer and sequence compression. We perform comprehensive experiments using passage and document retrieval datasets from MSMARCO, along with all public datasets from BEIR. In our experiments, Matryoshka Re-Ranker substantially outperforms existing methods, while effectively preserving its superior performance across various compression forms and application scenarios. We have publicly released our method at this https://github.com/FlagOpen/FlagEmbedding repo.

Dense retrieval calls for discriminative embeddings to represent the semantic relationship between query and document.It may benefit from the using of large language models (LLMs), given LLMs' strong capability on semantic understanding.However, the LLMs are learned by auto-regression, whose working mechanism is completely different from representing whole text as one discriminative embedding.Thus, it is imperative to study how to adapt LLMs properly so that they can be effectively initialized as the backbone encoder for dense retrieval.In this paper, we propose a novel approach, called Llama2Vec, which performs unsupervised adaptation of LLM for its dense retrieval application.Llama2Vec consists of two pretext tasks: EBAE (Embedding-Based Auto-Encoding) and EBAR (Embedding-Based Auto-Regression), where the LLM is prompted to reconstruct the input sentence and predict the next sentence based on its text embeddings.Llama2Vec is simple, lightweight, but highly effective.It is used to adapt LLaMA-2-7B on the Wikipedia corpus.With a moderate steps of adaptation, it substantially improves the model's fine-tuned performances on a variety of dense retrieval benchmarks.Notably, it results in the new state-of-the-art performances on popular benchmarks, such as passage and document retrieval on MSMARCO, and zero-shot retrieval on BEIR.The model and source code will be made publicly available to facilitate the future research.Our model is available at https://github.com/FlagOpen/FlagEmbedding.

Retrieval augmentation is a promising approach to handle long-context language modeling.However, the existing retrieval methods usually work with the chunked context, which is prone to inferior quality of semantic representation and incomplete retrieval of useful information.In this work, we propose a new method for the retrieval augmentation of longcontext language modeling, called Landmark Embedding.Our method is characterized by threefold technical contributions.Firstly, we introduce a chunking-free architecture, which keeps the long context coherent such that highquality embeddings can be generated for the fine-grained units within the context.Secondly, we present a position-aware objective function, which prioritizes the ultimate boundary for a consecutive span of information.By learning to discriminate such a special position, the useful information can be comprehensively retrieved for the query.Thirdly, we design a novel multistage learning algorithm, which makes the best use of readily available data and synthetic data for cost-effective training of the landmark embedding.In our experimental study, landmark embedding is able to substantially improve the performance for both LLaMA-2 and ChatGPT in a variety of long-context tasks; meanwhile, it also outperforms the existing retrieval methods with a notable advantage.Our model and code will be made publicly available 1 .

LLMs confront inherent limitations in terms of its knowledge, memory, and action.The retrieval augmentation stands as a vital mechanism to address these limitations, which brings in useful information from external sources to augment the LLM.However, existing retrieval methods encounter two pressing issues.On one hand, the general retrievers are not properly optimized for retrieval augmentation hence exhibit limited effectiveness; on the other hand, the task-specific retrievers excel in the targeted retrieval augmentation scenario, while lack the versatility to handle diverse scenarios.In this work, we propose LLM-Embedder for the unified support of diverse retrieval augmentation scenarios.Our method presents three technical contributions.Firstly, we introduce a new reward formulation, namely rank-aware reward.It exploits the ranking position of the desired output among N sampled outputs from the LLM, which leads to fine-grained and robust computation of reward from the LLM's feedback.Secondly, we design a novel distillation objective, called graded distillation.It incorporates both the absolute value and the relative order of the reward for more sufficient utilization of the LLM's feedback.Thirdly, we systematically optimize the multi-task learning, which effectively unifies the multiple retrieval functionalities into one model.In our experiment, LLM-Embedder notably improves the LLM's performances in various downstream tasks, and outperforms both general and task-specific retrievers with a substantial advantage.Our data, code, and model have been released at https: //github.com/FlagOpen/FlagEmbedding.

Multi-modal retrieval becomes increasingly popular in practice.However, the existing retrievers are mostly text-oriented, which lack the capability to process visual information.Despite the presence of vision-language models like CLIP, the current methods are severely limited in representing the text-only and imageonly data.In this work, we present a new embedding model VISTA for universal multimodal retrieval.Our work brings forth threefold technical contributions.Firstly, we introduce a flexible architecture which extends a powerful text encoder with the image understanding capability by introducing visual token embeddings.Secondly, we develop two data generation strategies, which bring highquality composed image-text to facilitate the training of the embedding model.Thirdly, we introduce a multi-stage training algorithm, which first aligns the visual token embedding with the text encoder using massive weakly labeled data, and then develops multi-modal representation capability using the generated composed image-text data.In our experiments, VISTA achieves superior performances across a variety of multi-modal retrieval tasks in both zero-shot and supervised settings.Our model, data, and source code are available at https://github.com/FlagOpen/FlagEmbedding.

Pre-trained language models like BERT and T5 serve as crucial backbone encoders for dense retrieval.However, these models often exhibit limited generalization capabilities and face challenges in improving in-domain accuracy.Recent research has explored using large language models (LLMs) as retrievers, achieving state-of-the-art performance across various tasks.Despite these advancements, the specific benefits of LLMs over traditional retrievers and the impact of different LLM configurations-such as parameter sizes, pre-training duration, and alignment processes-on retrieval tasks remain unclear.In this work, we conduct a comprehensive empirical study on six key dimensions of dense retrieval capabilities, including in-domain accuracy, data efficiency, zero-shot generalization, lengthy retrieval, instruction-based retrieval, and multi-task learning.We evaluate over 15 different backbone LLMs and non-LLMs.Our findings reveal that larger models and extensive pre-training consistently enhance in-domain accuracy and data efficiency.Additionally, larger models demonstrate significant potential in zero-shot generalization, lengthy retrieval, instruction-based retrieval, and multi-task learning.These results underscore the advantages of LLMs as versatile and effective backbone encoders in dense retrieval, providing valuable insights for future research and development in this field.

Towards energy-efficient artificial intelligence similar to the human brain, the bio-inspired spiking neural networks (SNNs) have advantages of biological plausibility, event-driven sparsity, and binary activation. Recently, large-scale language models exhibit promising generalization capability, making it a valuable issue to explore more general spike-driven models. However, the binary spikes in existing SNNs fail to encode adequate semantic information, placing technological challenges for generalization. This work proposes the first fully spiking mechanism for general language tasks, including both discriminative and generative ones. Different from previous spikes with 0,1 levels, we propose a more general spike formulation with bi-directional, elastic amplitude, and elastic frequency encoding, while still maintaining the addition nature of SNNs. In a single time step, the spike is enhanced by direction and amplitude information; in spike frequency, a strategy to control spike firing rate is well designed. We plug this elastic bi-spiking mechanism in language modeling, named SpikeLM. It is the first time to handle general language tasks with fully spike-driven models, which achieve much higher accuracy than previously possible. SpikeLM also greatly bridges the performance gap between SNNs and ANNs in language modeling. Our code is available at https://github.com/Xingrun-Xing/SpikeLM.

We introduce C-Pack, a package of resources that significantly advances the field of general text embeddings for Chinese. C-Pack includes three critical resources. 1) C-MTP is a massive training dataset for text embedding, which is based on the curation of vast unlabeled corpora and the integration of high-quality labeled corpora. 2) C-MTEB is a comprehensive benchmark for Chinese text embeddings covering 6 tasks and 35 datasets. 3) BGE is a family of embedding models covering multiple sizes. Our models outperform all prior Chinese text embeddings on C-MTEB by more than +10% upon the time of the release. We also integrate and optimize the entire suite of training methods for BGE. Along with our resources on general Chinese embedding, we release our data and models for English text embeddings. The English models also achieve state-of-the-art performance on the MTEB benchmark; meanwhile, our released English data is 2 times larger than the Chinese data. Both Chinese and English datasets are the largest public release of training data for text embeddings. All these resources are made publicly available at https://github.com/FlagOpen/FlagEmbedding.

To better support information retrieval tasks such as web search and open-domain question answering, growing effort is made to develop retrieval-oriented language models, e.g., RetroMAE (Xiao et al., 2022b) and many others (Gao and Callan, 2021; Wang et al., 2021a).Most of the existing works focus on improving the semantic representation capability for the contextualized embedding of the [CLS] token.However, recent study shows that the ordinary tokens besides [CLS] may provide extra information, which help to produce a better representation effect (Lin et al., 2022).As such, it's necessary to extend the current methods where all contextualized embeddings can be jointly pre-trained for the retrieval tasks.In this work, we propose a novel pre-training method called Duplex Masked Auto-Encoder, a.k.a.DupMAE.It is designed to improve the quality of semantic representation where all contextualized embeddings of the pre-trained model can be leveraged.It takes advantage of two complementary auto-encoding tasks: one reconstructs the input sentence with the [CLS] embedding; the other one predicts the bagof-words feature of the input sentence with the ordinary tokens' embeddings.The two tasks are jointly conducted to train a unified encoder, where the whole contextualized embeddings are aggregated in a compact way to produce the final semantic representation.DupMAE is simple but empirically competitive: it substantially improves the pre-trained model's representation capability and transferability, where superior retrieval performances can be achieved on popular benchmarks, like MS MARCO and BEIR.Our

Diversity is a vital factor in recommendation systems.Improving the diversity in recommendations helps broaden users' horizons, bring good user experience and promote the enterprises' sales. In the past years, many efforts have been devoted to optimizing the diversity in the matching stage and the re-ranking stage of the recommendation system, but few in the ranking stage. The ranking stage is the intermediate stage of the recommendation system. Improving the diversity of the ranking stage can preserve the diversity of the matching stage, and provide a more diversified list for the re-ranking stage. Besides, the ranking models are able to achieve a better balance between accuracy and diversity. In this paper, we aim to improve the diversity in the ranking stage. To address the diversity challenges posed by the pointwise ranking model and biased user interaction history, we propose a Diversity-aware Deep Ranking Network by carefully designing two diversity-aware components that are diversity-aware listwise information fusion and balanced weighting loss. We conduct both offline and online experiments, and the results demonstrate that our proposed model effectively improves the recommendation diversity in the ranking stage while maintaining the accuracy. Moreover, the new model achieves 1.27%, 2.30% and 1.98% improvements in VBR, GMV and Coverage in Meituan, one of the world's largest E-commerce platforms.

Inverted file structure is a common technique for accelerating dense retrieval.It clusters documents based on their embeddings; during searching, it probes nearby clusters w.r.t. an input query and only evaluates documents within them by subsequent codecs, thus avoiding the expensive cost of exhaustive traversal.However, the clustering is always lossy, which results in the miss of relevant documents in the probed clusters and hence degrades retrieval quality.In contrast, lexical matching, such as overlaps of salient terms, tends to be strong feature for identifying relevant documents.In this work, we present the Hybrid Inverted Index (HI 2 ), where the embedding clusters and salient terms work collaboratively to accelerate dense retrieval.To make best of both effectiveness and efficiency, we devise a cluster selector and a term selector, to construct compact inverted lists and efficiently searching through them.Moreover, we leverage simple unsupervised algorithms as well as end-to-end knowledge distillation to learn these two modules, with the latter further boosting the effectiveness.Based on comprehensive experiments on popular retrieval benchmarks, we verify that clusters and terms indeed complement each other, enabling HI 2 to achieve lossless retrieval quality with competitive efficiency across various index settings.

Vector quantization is one of the critical techniques which enables dense retrieval for realtime applications. The recent study shows that vanilla vector quantization methods, like those implemented by FAISS [8], are lossy and prone to limited retrieval performances when large acceleration ratios are needed [14, 16, 18]. Besides, there have also been multiple algorithms which make the retriever and VQ better collaborated to alleviate such a loss. On top of these progresses, we develop LibVQ, which optimizes vector quantization for efficient dense retrieval. Our toolkit is highlighted for three advantages. 1. Effectiveness. The retrieval quality can be substantially improved over the vanilla implementations of VQ. 2. Simplicity. The optimization can be conducted in a lowcode fashion, and the optimization results can be easily loaded to ANN indexes to support downstream applications. 3. Universality. The optimization is agnostic to the embedding's learning process, and may accommodate different input conditions and ANN back-ends with little modification of the workflow. LibVQ may also support rich applications beyond dense retrieval, e.g., embedding compression, topic modeling, and de-duplication. In this demo, we provide comprehensive hand-on examples and evaluations for LibVQ. The toolkit is publicly released at: https://github.com/staoxiao/LibVQ/tree/demo.

Despite pre-training's progress in many important NLP tasks, it remains to explore effective pre-training strategies for dense retrieval.In this paper, we propose RetroMAE, a new retrieval oriented pre-training paradigm based on Masked Auto-Encoder (MAE).RetroMAE is highlighted by three critical designs.1) A novel MAE workflow, where the input sentence is polluted for encoder and decoder with different masks.The sentence embedding is generated from the encoder's masked input; then, the original sentence is recovered based on the sentence embedding and the decoder's masked input via masked language modeling.2) Asymmetric model structure, with a full-scale BERT like transformer as encoder, and a one-layer transformer as decoder.3) Asymmetric masking ratios, with a moderate ratio for encoder: 15∼30%, and an aggressive ratio for decoder: 50∼70%.Our framework is simple to realize and empirically competitive: the pre-trained models dramatically improve the SOTA performances on a wide range of dense retrieval benchmarks, like BEIR and MS MARCO.The source code and pre-trained models are made publicly available at https://github.com/staoxiao/RetroMAEso as to inspire more interesting research.

News recommendation calls for deep insights of news articles' underlying semantics. Therefore, pretrained language models (PLMs), like BERT and RoBERTa, may substantially contribute to the recommendation quality. However, it's extremely challenging to have news recommenders trained together with such big models: the learning of news recommenders requires intensive news encoding operations, whose cost is prohibitive if PLMs are used as the news encoder. In this paper, we propose a novel framework, SpeedyFeed, which efficiently trains PLMs-based news recommenders of superior quality. SpeedyFeed is highlighted for its light-weight encoding pipeline, which gives rise to three major advantages. Firstly, it makes the intermediate results fully reusable for the training workflow, which removes most of the repetitive but redundant encoding operations. Secondly, it improves the data efficiency of the training workflow, where non-informative data can be eliminated from encoding. Thirdly, it further saves the cost by leveraging simplified news encoding and compact news representation.

Embedding based retrieval (EBR) is a fundamental building block in many web applications. However, EBR in sponsored search is distinguished from other generic scenarios and technically challenging due to the need of serving multiple retrieval purposes: firstly, it has to retrieve high-relevance ads, which may exactly serve user's search intent; secondly, it needs to retrieve high-CTR ads so as to maximize the overall user clicks. In this paper, we present a novel representation learning framework Uni-Retriever developed for Bing Search, which unifies two different training modes knowledge distillation and contrastive learning to realize both required objectives. On one hand, the capability of making high-relevance retrieval is established by distilling knowledge from the "relevance teacher model''. On the other hand, the capability of making high-CTR retrieval is optimized by learning to discriminate user's clicked ads from the entire corpus. The two training modes are jointly performed as a multi-objective learning process, such that the ads of high relevance and CTR can be favored by the generated embeddings. Besides the learning strategy, we also elaborate our solution for EBR serving pipeline built upon the substantially optimized DiskANN, where massive-scale EBR can be performed with competitive time and memory efficiency, and accomplished in high-quality. We make comprehensive offline and online experiments to evaluate the proposed techniques, whose findings may provide useful insights for the future development of EBR systems. Uni-Retriever has been mainstreamed as the major retrieval path in Bing's production thanks to the notable improvements on the representation and EBR serving quality.

Vector quantization (VQ) based ANN indexes, such as Inverted File System (IVF) and Product Quantization (PQ), have been widely applied to embedding based document retrieval thanks to the competitive time and memory efficiency. Originally, VQ is learned to minimize the reconstruction loss, i.e., the distortions between the original dense embeddings and the reconstructed embeddings after quantization. Unfortunately, such an objective is inconsistent with the goal of selecting ground-truth documents for the input query, which may cause severe loss of retrieval quality. Recent works identify such a defect, and propose to minimize the retrieval loss through contrastive learning. However, these methods intensively rely on queries with ground-truth documents, whose performance is limited by the insufficiency of labeled data. In this paper, we propose Distill-VQ, which unifies the learning of IVF and PQ within a knowledge distillation framework. In Distill-VQ, the dense embeddings are leveraged as "teachers'', which predict the query's relevance to the sampled documents. The VQ modules are treated as the "students'', which are learned to reproduce the predicted relevance, such that the reconstructed embeddings may fully preserve the retrieval result of the dense embeddings. By doing so, Distill-VQ is able to derive substantial training signals from the massive unlabeled data, which significantly contributes to the retrieval quality. We perform comprehensive explorations for the optimal conduct of knowledge distillation, which may provide useful insights for the learning of VQ based ANN index. We also experimentally show that the labeled data is no longer a necessity for high-quality vector quantization, which indicates Distill-VQ's strong applicability in practice. The evaluations are performed on MS MARCO and Natural Questions benchmarks, where Distill-VQ notably outperforms the SOTA VQ methods in Recall and MRR. Our code is avaliable at https://github.com/staoxiao/LibVQ.

Recommender system is playing an increasingly important role in online news platforms nowadays. Recently, there is a growing demand for applying reinforcement learning (RL) algorithms to news recommendation aiming to maximize long-term and/or non-differentiable objectives. However, without an interactive simulated environment, it is extremely costly to develop powerful RL agents for news recommendation. In this paper, we build a user simulator, namely MINDSim, for news recommendation. Targeting at new user generation and corresponding behavior simulation, we first construct a hidden space for users using a generative adversarial network, so that new users can be generated by sampling from this hidden space. To capture complex and fast user interest drifts over time, we adopt an encoder-decoder architecture, which takes the clicked news during the simulation as input and outputs the new user interests for the next period of time. Finally, we build the MINDSim simulator using MIcrosoft News Dataset (MIND), and extensive experimental results on this large-scale real-world dataset demonstrate that MINDSim can simulate the behaviors of real users with high quality.

Ad-hoc search calls for the selection of appropriate answers from a massive-scale corpus. Nowadays, the embedding-based retrieval (EBR) becomes a promising solution, where deep learning based document representation and ANN search techniques are allied to handle this task. However, a major challenge is that the ANN index can be too large to fit into memory, given the considerable size of answer corpus. In this work, we tackle this problem with Bi-Granular Document Representation, where the lightweight sparse embeddings are indexed and standby in memory for coarse-grained candidate search, and the heavyweight dense embeddings are hosted in disk for fine-grained post verification. For the best of retrieval accuracy, a Progressive Optimization framework is designed. The sparse embeddings are learned ahead for high-quality search of candidates. Conditioned on the candidate distribution induced by the sparse embeddings, the dense embeddings are continuously learned to optimize the discrimination of ground-truth from the shortlisted candidates. Besides, two techniques: the contrastive quantization and the locality-centric sampling are introduced for the learning of sparse and dense embeddings, which substantially contribute to their performances. Thanks to the above features, our method effectively handles massive-scale EBR with strong advantages in accuracy: with up to recall gain on million-scale corpus, and up to recall gain on billion-scale corpus. Besides, Our method is applied to a major sponsored search platform with substantial gains on revenue (), Recall () and CTR (). Our code is available at https://github.com/microsoft/BiDR.

Product quantization (PQ) is a widely used technique for ad-hoc retrieval.Recent studies propose supervised PQ, where the embedding and quantization models can be jointly trained with supervised learning.However, there is a lack of appropriate formulation of the joint training objective; thus, the improvements over previous non-supervised baselines are limited in reality.In this work, we propose the Matching-oriented Product Quantization (MoPQ), where a novel objective Multinoulli Contrastive Loss (MCL) is formulated.With the minimization of MCL, we are able to maximize the matching probability of query and ground-truth key, which contributes to the optimal retrieval accuracy.Given that the exact computation of MCL is intractable due to the demand of vast contrastive samples, we further propose the Differentiable Cross-device Sampling (DCS), which significantly augments the contrastive samples for precise approximation of MCL.We conduct extensive experimental studies on four realworld datasets, whose results verify the effectiveness of MoPQ.The code is available at https://github.com/microsoft/MoPQ.1. We adopt the Asymmetric Distance Computation (ADC) (Jégou et al., 2011), where only keys need to be quantized.

The representation learning on textual graph is to generate low-dimensional embeddings for the nodes based on the individual textual features and the neighbourhood information. Recent breakthroughs on pretrained language models and graph neural networks push forward the development of corresponding techniques. The existing works mainly rely on the cascaded model architecture: the textual features of nodes are independently encoded by language models at first; the textual embeddings are aggregated by graph neural networks afterwards. However, the above architecture is limited due to the independent modeling of textual features. In this work, we propose GraphFormers, where layerwise GNN components are nested alongside the transformer blocks of language models. With the proposed architecture, the text encoding and the graph aggregation are fused into an iterative workflow, making each node's semantic accurately comprehended from the global perspective. In addition, a progressive learning strategy is introduced, where the model is successively trained on manipulated data and original data to reinforce its capability of integrating information on graph. Extensive evaluations are conducted on three large-scale benchmark datasets, where GraphFormers outperform the SOTA baselines with comparable running efficiency. The source code is released at https://github.com/microsoft/GraphFormers .