Jian Jiao

GLM-5 Team, Aohan Zeng, Xin Lv et al. · tsinghua

We present GLM-5, a next-generation foundation model designed to transition the paradigm of vibe coding to agentic engineering. Building upon the agentic, reasoning, and coding (ARC) capabilities of its predecessor, GLM-5 adopts DSA to significantly reduce training and inference costs while maintaining long-context fidelity. To advance model alignment and autonomy, we implement a new asynchronous reinforcement learning infrastructure that drastically improves post-training efficiency by decoupling generation from training. Furthermore, we propose novel asynchronous agent RL algorithms that further improve RL quality, enabling the model to learn from complex, long-horizon interactions more effectively. Through these innovations, GLM-5 achieves state-of-the-art performance on major open benchmarks. Most critically, GLM-5 demonstrates unprecedented capability in real-world coding tasks, surpassing previous baselines in handling end-to-end software engineering challenges. Code, models, and more information are available at https://github.com/zai-org/GLM-5.

Zhenghao Lin, Yeyun Gong, Xiao Liu et al. · microsoft-research

Knowledge distillation is an effective way to transfer knowledge from a strong teacher to an efficient student model. Ideally, we expect the better the teacher is, the better the student. However, this expectation does not always come true. It is common that a better teacher model results in a bad student via distillation due to the nonnegligible gap between teacher and student. To bridge the gap, we propose PROD, a PROgressive Distillation method, for dense retrieval. PROD consists of a teacher progressive distillation and a data progressive distillation to gradually improve the student. We conduct extensive experiments on five widely-used benchmarks, MS MARCO Passage, TREC Passage 19, TREC Document 19, MS MARCO Document and Natural Questions, where PROD achieves the state-of-the-art within the distillation methods for dense retrieval. The code and models will be released.

Simiao Zuo, Xiaodong Liu, Jian Jiao et al. · microsoft-research

Transformer models have achieved superior performance in various natural language processing tasks. However, the quadratic computational cost of the attention mechanism limits its practicality for long sequences. There are existing attention variants that improve the computational efficiency, but they have limited ability to effectively compute global information. In parallel to Transformer models, state space models (SSMs) are tailored for long sequences, but they are not flexible enough to capture complicated local information. We propose SPADE, short for $\underline{\textbf{S}}$tate s$\underline{\textbf{P}}$ace $\underline{\textbf{A}}$ugmente$\underline{\textbf{D}}$ Transform$\underline{\textbf{E}}$r. Specifically, we augment a SSM into the bottom layer of SPADE, and we employ efficient local attention methods for the other layers. The SSM augments global information, which complements the lack of long-range dependency issue in local attention methods. Experimental results on the Long Range Arena benchmark and language modeling tasks demonstrate the effectiveness of the proposed method. To further demonstrate the scalability of SPADE, we pre-train large encoder-decoder models and present fine-tuning results on natural language understanding and natural language generation tasks.

Weizhen Qi, Yeyun Gong, Yelong Shen et al. · microsoft-research

Non-Autoregressive generation is a sequence generation paradigm, which removes the dependency between target tokens. It could efficiently reduce the text generation latency with parallel decoding in place of token-by-token sequential decoding. However, due to the known multi-modality problem, Non-Autoregressive (NAR) models significantly under-perform Auto-regressive (AR) models on various language generation tasks. Among the NAR models, BANG is the first large-scale pre-training model on English un-labeled raw text corpus. It considers different generation paradigms as its pre-training tasks including Auto-regressive (AR), Non-Autoregressive (NAR), and semi-Non-Autoregressive (semi-NAR) information flow with multi-stream strategy. It achieves state-of-the-art performance without any distillation techniques. However, AR distillation has been shown to be a very effective solution for improving NAR performance. In this paper, we propose a novel self-paced mixed distillation method to further improve the generation quality of BANG. Firstly, we propose the mixed distillation strategy based on the AR stream knowledge. Secondly, we encourage the model to focus on the samples with the same modality by self-paced learning. The proposed self-paced mixed distillation algorithm improves the generation quality and has no influence on the inference latency. We carry out extensive experiments on summarization and question generation tasks to validate the effectiveness. To further illustrate the commercial value of our approach, we conduct experiments on three generation tasks in real-world advertisements applications. Experimental results on commercial data show the effectiveness of the proposed model. Compared with BANG, it achieves significant BLEU score improvement. On the other hand, compared with auto-regressive generation method, it achieves more than 7x speedup.

Xingwei He, Zhenghao Lin, Yeyun Gong et al.

Many natural language processing (NLP) tasks rely on labeled data to train machine learning models with high performance. However, data annotation is time-consuming and expensive, especially when the task involves a large amount of data or requires specialized domains. Recently, GPT-3.5 series models have demonstrated remarkable few-shot and zero-shot ability across various NLP tasks. In this paper, we first claim that large language models (LLMs), such as GPT-3.5, can serve as an excellent crowdsourced annotator when provided with sufficient guidance and demonstrated examples. Accordingly, we propose AnnoLLM, an annotation system powered by LLMs, which adopts a two-step approach, explain-then-annotate. Concretely, we first prompt LLMs to provide explanations for why the specific ground truth answer/label was assigned for a given example. Then, we construct the few-shot chain-of-thought prompt with the self-generated explanation and employ it to annotate the unlabeled data with LLMs. Our experiment results on three tasks, including user input and keyword relevance assessment, BoolQ, and WiC, demonstrate that AnnoLLM surpasses or performs on par with crowdsourced annotators. Furthermore, we build the first conversation-based information retrieval dataset employing AnnoLLM. This dataset is designed to facilitate the development of retrieval models capable of retrieving pertinent documents for conversational text. Human evaluation has validated the dataset's high quality.

Sanxing Chen, Hao Cheng, Xiaodong Liu et al. · microsoft-research

Learning transferable representation of knowledge graphs (KGs) is challenging due to the heterogeneous, multi-relational nature of graph structures. Inspired by Transformer-based pretrained language models' success on learning transferable representation for texts, we introduce a novel inductive KG representation model (iHT) for KG completion by large-scale pre-training. iHT consists of a entity encoder (e.g., BERT) and a neighbor-aware relational scoring function both parameterized by Transformers. We first pre-train iHT on a large KG dataset, Wikidata5M. Our approach achieves new state-of-the-art results on matched evaluations, with a relative improvement of more than 25% in mean reciprocal rank over previous SOTA models. When further fine-tuned on smaller KGs with either entity and relational shifts, pre-trained iHT representations are shown to be transferable, significantly improving the performance on FB15K-237 and WN18RR.

Hong Sun, Xue Li, Yinchuan Xu et al. · microsoft-research

This paper presents AutoHint, a novel framework for automatic prompt engineering and optimization for Large Language Models (LLM). While LLMs have demonstrated remarkable ability in achieving high-quality annotation in various tasks, the key to applying this ability to specific tasks lies in developing high-quality prompts. Thus we propose a framework to inherit the merits of both in-context learning and zero-shot learning by incorporating enriched instructions derived from input-output demonstrations to optimize original prompt. We refer to the enrichment as the hint and propose a framework to automatically generate the hint from labeled data. More concretely, starting from an initial prompt, our method first instructs a LLM to deduce new hints for selected samples from incorrect predictions, and then summarizes from per-sample hints and adds the results back to the initial prompt to form a new, enriched instruction. The proposed method is evaluated on the BIG-Bench Instruction Induction dataset for both zero-shot and few-short prompts, where experiments demonstrate our method is able to significantly boost accuracy for multiple tasks.

Kunal Dahiya, Nilesh Gupta, Deepak Saini et al.

Extreme Classification (XC) seeks to tag data points with the most relevant subset of labels from an extremely large label set. Performing deep XC with dense, learnt representations for data points and labels has attracted much attention due to its superiority over earlier XC methods that used sparse, hand-crafted features. Negative mining techniques have emerged as a critical component of all deep XC methods that allow them to scale to millions of labels. However, despite recent advances, training deep XC models with large encoder architectures such as transformers remains challenging. This paper identifies that memory overheads of popular negative mining techniques often force mini-batch sizes to remain small and slow training down. In response, this paper introduces NGAME, a light-weight mini-batch creation technique that offers provably accurate in-batch negative samples. This allows training with larger mini-batches offering significantly faster convergence and higher accuracies than existing negative sampling techniques. NGAME was found to be up to 16% more accurate than state-of-the-art methods on a wide array of benchmark datasets for extreme classification, as well as 3% more accurate at retrieving search engine queries in response to a user webpage visit to show personalized ads. In live A/B tests on a popular search engine, NGAME yielded up to 23% gains in click-through-rates.

Xingwei He, Yeyun Gong, A-Long Jin et al.

Commonsense generation aims to generate a realistic sentence describing a daily scene under the given concepts, which is very challenging, since it requires models to have relational reasoning and compositional generalization capabilities. Previous work focuses on retrieving prototype sentences for the provided concepts to assist generation. They first use a sparse retriever to retrieve candidate sentences, then re-rank the candidates with a ranker. However, the candidates returned by their ranker may not be the most relevant sentences, since the ranker treats all candidates equally without considering their relevance to the reference sentences of the given concepts. Another problem is that re-ranking is very expensive, but only using retrievers will seriously degrade the performance of their generation models. To solve these problems, we propose the metric distillation rule to distill knowledge from the metric (e.g., BLEU) to the ranker. We further transfer the critical knowledge summarized by the distilled ranker to the retriever. In this way, the relevance scores of candidate sentences predicted by the ranker and retriever will be more consistent with their quality measured by the metric. Experimental results on the CommonGen benchmark verify the effectiveness of our proposed method: (1) Our generation model with the distilled ranker achieves a new state-of-the-art result. (2) Our generation model with the distilled retriever even surpasses the previous SOTA.

Xingwei He, Yeyun Gong, A-Long Jin et al.

The dual-encoder has become the de facto architecture for dense retrieval. Typically, it computes the latent representations of the query and document independently, thus failing to fully capture the interactions between the query and document. To alleviate this, recent research has focused on obtaining query-informed document representations. During training, it expands the document with a real query, but during inference, it replaces the real query with a generated one. This inconsistency between training and inference causes the dense retrieval model to prioritize query information while disregarding the document when computing the document representation. Consequently, it performs even worse than the vanilla dense retrieval model because its performance heavily relies on the relevance between the generated queries and the real query.In this paper, we propose a curriculum sampling strategy that utilizes pseudo queries during training and progressively enhances the relevance between the generated query and the real query. By doing so, the retrieval model learns to extend its attention from the document alone to both the document and query, resulting in high-quality query-informed document representations. Experimental results on both in-domain and out-of-domain datasets demonstrate that our approach outperforms previous dense retrieval models.

Jatin Prakash, Anirudh Buvanesh, Bishal Santra et al.

Extreme Classification (XC) aims to map a query to the most relevant documents from a very large document set. XC algorithms used in real-world applications learn this mapping from datasets curated from implicit feedback, such as user clicks. However, these datasets inevitably suffer from missing labels. In this work, we observe that systematic missing labels lead to missing knowledge, which is critical for accurately modelling relevance between queries and documents. We formally show that this absence of knowledge cannot be recovered using existing methods such as propensity weighting and data imputation strategies that solely rely on the training dataset. While LLMs provide an attractive solution to augment the missing knowledge, leveraging them in applications with low latency requirements and large document sets is challenging. To incorporate missing knowledge at scale, we propose SKIM (Scalable Knowledge Infusion for Missing Labels), an algorithm that leverages a combination of small LM and abundant unstructured meta-data to effectively mitigate the missing label problem. We show the efficacy of our method on large-scale public datasets through exhaustive unbiased evaluation ranging from human annotations to simulations inspired from industrial settings. SKIM outperforms existing methods on Recall@100 by more than 10 absolute points. Additionally, SKIM scales to proprietary query-ad retrieval datasets containing 10 million documents, outperforming contemporary methods by 12% in offline evaluation and increased ad click-yield by 1.23% in an online A/B test conducted on a popular search engine. We release our code, prompts, trained XC models and finetuned SLMs at: https://github.com/bicycleman15/skim

Xinyu Hu, Pengfei Tang, Simiao Zuo et al.

Large language models (LLMs) have made impressive progress in natural language processing. These models rely on proper human instructions (or prompts) to generate suitable responses. However, the potential of LLMs are not fully harnessed by commonly-used prompting methods: many human-in-the-loop algorithms employ ad-hoc procedures for prompt selection; while auto prompt generation approaches are essentially searching all possible prompts randomly and inefficiently. We propose Evoke, an automatic prompt refinement framework. In Evoke, there are two instances of a same LLM: one as a reviewer (LLM-Reviewer), it scores the current prompt; the other as an author (LLM-Author), it edits the prompt by considering the edit history and the reviewer's feedback. Such an author-reviewer feedback loop ensures that the prompt is refined in each iteration. We further aggregate a data selection approach to Evoke, where only the hard samples are exposed to the LLM. The hard samples are more important because the LLM can develop deeper understanding of the tasks out of them, while the model may already know how to solve the easier cases. Experimental results show that Evoke significantly outperforms existing methods. For instance, in the challenging task of logical fallacy detection, Evoke scores above 80, while all other baseline methods struggle to reach 20.

GLM-4. 5 Team, Aohan Zeng, Xin Lv et al.

We present GLM-4.5, an open-source Mixture-of-Experts (MoE) large language model with 355B total parameters and 32B activated parameters, featuring a hybrid reasoning method that supports both thinking and direct response modes. Through multi-stage training on 23T tokens and comprehensive post-training with expert model iteration and reinforcement learning, GLM-4.5 achieves strong performance across agentic, reasoning, and coding (ARC) tasks, scoring 70.1% on TAU-Bench, 91.0% on AIME 24, and 64.2% on SWE-bench Verified. With much fewer parameters than several competitors, GLM-4.5 ranks 3rd overall among all evaluated models and 2nd on agentic benchmarks. We release both GLM-4.5 (355B parameters) and a compact version, GLM-4.5-Air (106B parameters), to advance research in reasoning and agentic AI systems. Code, models, and more information are available at https://github.com/zai-org/GLM-4.5.

Yaoxiang Wang, Simiao Zuo, Qingguo Hu et al.

Embedding models are pivotal in industrial information retrieval systems like search and advertising. However, existing pretrained models often exhibit fixed architectures and embedding dimensionalities, posing significant challenges when adapting them to diverse deployment scenarios with varying business-driven constraints. A common practice involves fine-tuning with partial parameter initialization from larger pretrained models for resource-constrained tasks. This method is often suboptimal as the misalignment between pretraining and downstream usage prevents full realization of pretraining benefits. To address this limitation, we introduce m3BERT: a Modern, Multi-lingual, Matryoshka Bidirectional Encoder, which features a novel pretraining strategy that jointly optimizes representations across both transformer layers and multiple embedding dimensions. This enables a single model to be tailored to varied resource and accuracy targets while maintaining consistency with pretraining. Incorporating recent architectural improvements, m3BERT uses a three-stage pretraining: monolingual pretraining, multilingual adaptation to serve diverse user bases, and crucial continual pretraining on a massive web domain corpus to enhance utility in commercial retrieval. m3BERT significantly outperforms state-of-the-art embedding models in Bing-Click, a large-scale industrial retrieval dataset, showcasing its practical versatility as an efficient foundation for resource-aware industrial retrieval systems. Further experiments on public datasets also confirm the general effectiveness of our multigranular Matryoshka pretraining strategy.

Amit Sharma, Hua Li, Jian Jiao

Given an unexpected change in the output metric of a large-scale system, it is important to answer why the change occurred: which inputs caused the change in metric? A key component of such an attribution question is estimating the counterfactual: the (hypothetical) change in the system metric due to a specified change in a single input. However, due to inherent stochasticity and complex interactions between parts of the system, it is difficult to model an output metric directly. We utilize the computational structure of a system to break up the modelling task into sub-parts, such that each sub-part corresponds to a more stable mechanism that can be modelled accurately over time. Using the system's structure also helps to view the metric as a computation over a structural causal model (SCM), thus providing a principled way to estimate counterfactuals. Specifically, we propose a method to estimate counterfactuals using time-series predictive models and construct an attribution score, CF-Shapley, that is consistent with desirable axioms for attributing an observed change in the output metric. Unlike past work on causal shapley values, our proposed method can attribute a single observed change in output (rather than a population-level effect) and thus provides more accurate attribution scores when evaluated on simulated datasets. As a real-world application, we analyze a query-ad matching system with the goal of attributing observed change in a metric for ad matching density. Attribution scores explain how query volume and ad demand from different query categories affect the ad matching density, leading to actionable insights and uncovering the role of external events (e.g., "Cheetah Day") in driving the matching density.

Victor Ye Dong, Kuan-Yun Lee, Jiamei Shuai et al.

We present \emph{Greedy Information Projection} (\textsc{GIP}), a principled framework for choosing training examples for large language model fine-tuning. \textsc{GIP} casts selection as maximizing mutual information between a subset of examples and task-specific query signals, which may originate from LLM quality judgments, metadata, or other sources. The framework involves optimizing a closed-form mutual information objective defined using both data and query embeddings, naturally balancing {\it quality} and {\it diversity}. Optimizing this score is equivalent to maximizing the projection of the query embedding matrix onto the span of the selected data, which provides a geometric explanation for the co-emergence of quality and diversity. Building on this view, we employ a fast greedy matching-pursuit procedure with efficient projection-based updates. On instruction-following and mathematical reasoning datasets, \textsc{GIP} selects small subsets that match full-data fine-tuning while using only a fraction of examples and compute, unifying quality-aware and diversity-aware selection for efficient fine-tuning.

Liang Luo, Yinbin Ma, Quanyu Zhu et al.

Recent GPU generations deliver significantly higher FLOPs using lower-precision arithmetic, such as FP8. While successfully applied to large language models (LLMs), its adoption in large recommendation models (LRMs) has been limited. This is because LRMs are numerically sensitive, dominated by small matrix multiplications (GEMMs) followed by normalization, and trained in communication-intensive environments. Applying FP8 directly to LRMs often degrades model quality and prolongs training time. These challenges are inherent to LRM workloads and cannot be resolved merely by introducing better FP8 kernels. Instead, a system-model co-design approach is needed to successfully integrate FP8. We present LoKA (Low-precision Kernel Applications), a framework that makes FP8 practical for LRMs through three principles: profile under realistic distributions to know where low precision is safe, co-design model components with hardware to expand where it is safe, and orchestrate across kernel libraries to maximize the gains. Concretely, LoKA Probe is a statistically grounded, online benchmarking method that learns activation and weight statistics, and quantifies per-layer errors. This process pinpoints safe and unsafe, fast and slow sites for FP8 adoption. LoKA Mods is a set of reusable model adaptations that improve both numerical stability and execution efficiency with FP8. LoKA Dispatch is a runtime that leverages the statistical insights from LoKA Probe to select the fastest FP8 kernel that satisfies the accuracy requirements.

Zhenghao Lin, Zhibin Gou, Yeyun Gong et al. · microsoft-research, tsinghua

Previous language model pre-training methods have uniformly applied a next-token prediction loss to all training tokens. Challenging this norm, we posit that "9l training". Our initial analysis examines token-level training dynamics of language model, revealing distinct loss patterns for different tokens. Leveraging these insights, we introduce a new language model called Rho-1. Unlike traditional LMs that learn to predict every next token in a corpus, Rho-1 employs Selective Language Modeling (SLM), which selectively trains on useful tokens that aligned with the desired distribution. This approach involves scoring pretraining tokens using a reference model, and then training the language model with a focused loss on tokens with higher scores. When continual pretraining on 15B OpenWebMath corpus, Rho-1 yields an absolute improvement in few-shot accuracy of up to 30% in 9 math tasks. After fine-tuning, Rho-1-1B and 7B achieved state-of-the-art results of 40.6% and 51.8% on MATH dataset, respectively - matching DeepSeekMath with only 3% of the pretraining tokens. Furthermore, when continual pretraining on 80B general tokens, Rho-1 achieves 6.8% average enhancement across 15 diverse tasks, increasing both efficiency and performance of the language model pre-training.

Tong Wu, Zhihao Fan, Xiao Liu et al.

Diffusion models have gained significant attention in the realm of image generation due to their exceptional performance. Their success has been recently expanded to text generation via generating all tokens within a sequence concurrently. However, natural language exhibits a far more pronounced sequential dependency in comparison to images, and the majority of existing language models are trained with a left-to-right auto-regressive approach. To account for the inherent sequential characteristic of natural language, we introduce Auto-Regressive Diffusion (AR-Diffusion). AR-Diffusion ensures that the generation of tokens on the right depends on the generated ones on the left, a mechanism achieved through employing a dynamic number of denoising steps that vary based on token position. This results in tokens on the left undergoing fewer denoising steps than those on the right, thereby enabling them to generate earlier and subsequently influence the generation of tokens on the right. In a series of experiments on various text generation tasks, including text summarization, machine translation, and common sense generation, AR-Diffusion clearly demonstrated its superiority over existing diffusion language models and that it can be $100\times\sim600\times$ faster when achieving comparable results. Our code is available at https://github.com/microsoft/ProphetNet/tree/master/AR-diffusion.

Ziheng Li, Shaohan Huang, Zihan Zhang et al.

Recent studies have shown that dual encoder models trained with the sentence-level translation ranking task are effective methods for cross-lingual sentence embedding. However, our research indicates that token-level alignment is also crucial in multilingual scenarios, which has not been fully explored previously. Based on our findings, we propose a dual-alignment pre-training (DAP) framework for cross-lingual sentence embedding that incorporates both sentence-level and token-level alignment. To achieve this, we introduce a novel representation translation learning (RTL) task, where the model learns to use one-side contextualized token representation to reconstruct its translation counterpart. This reconstruction objective encourages the model to embed translation information into the token representation. Compared to other token-level alignment methods such as translation language modeling, RTL is more suitable for dual encoder architectures and is computationally efficient. Extensive experiments on three sentence-level cross-lingual benchmarks demonstrate that our approach can significantly improve sentence embedding. Our code is available at https://github.com/ChillingDream/DAP.

Simiao Zuo, Xiaodong Liu, Jian Jiao et al.

Sparsely activated models (SAMs), such as Mixture-of-Experts (MoE), can easily scale to have outrageously large amounts of parameters without significant increase in computational cost. However, SAMs are reported to be parameter inefficient such that larger models do not always lead to better performance. While most on-going research focuses on improving SAMs models by exploring methods of routing inputs to experts, our analysis reveals that such research might not lead to the solution we expect, i.e., the commonly-used routing methods based on gating mechanisms do not work better than randomly routing inputs to experts. In this paper, we propose a new expert-based model, THOR (Transformer witH StOchastic ExpeRts). Unlike classic expert-based models, such as the Switch Transformer, experts in THOR are randomly activated for each input during training and inference. THOR models are trained using a consistency regularized loss, where experts learn not only from training data but also from other experts as teachers, such that all the experts make consistent predictions. We validate the effectiveness of THOR on machine translation tasks. Results show that THOR models are more parameter efficient in that they significantly outperform the Transformer and MoE models across various settings. For example, in multilingual translation, THOR outperforms the Switch Transformer by 2 BLEU scores, and obtains the same BLEU score as that of a state-of-the-art MoE model that is 18 times larger. Our code is publicly available at: https://github.com/microsoft/Stochastic-Mixture-of-Experts.

Dayiheng Liu, Yu Yan, Yeyun Gong et al.

Multi-task benchmarks such as GLUE and SuperGLUE have driven great progress of pretraining and transfer learning in Natural Language Processing (NLP). These benchmarks mostly focus on a range of Natural Language Understanding (NLU) tasks, without considering the Natural Language Generation (NLG) models. In this paper, we present the General Language Generation Evaluation (GLGE), a new multi-task benchmark for evaluating the generalization capabilities of NLG models across eight language generation tasks. For each task, we continue to design three subtasks in terms of task difficulty (GLGE-Easy, GLGE-Medium, and GLGE-Hard). This introduces 24 subtasks to comprehensively compare model performance. To encourage research on pretraining and transfer learning on NLG models, we make GLGE publicly available and build a leaderboard with strong baselines including MASS, BART, and ProphetNet (The source code and dataset are publicly available at https://github.com/microsoft/glge).

Qinyu Zhang, Liang Xu, Jianhao Huang et al.

Driven by the vision of ubiquitous connectivity and wireless intelligence, the evolution of ultra-dense constellation-based satellite-integrated Internet is underway, now taking preliminary shape. Nevertheless, the entrenched institutional silos and limited, nonrenewable heterogeneous network resources leave current satellite systems struggling to accommodate the escalating demands of next-generation intelligent applications. In this context, the distributed satellite information networks (DSIN), exemplified by the cohesive clustered satellites system, have emerged as an innovative architecture, bridging information gaps across diverse satellite systems, such as communication, navigation, and remote sensing, and establishing a unified, open information network paradigm to support resilient space information services. This survey first provides a profound discussion about innovative network architectures of DSIN, encompassing distributed regenerative satellite network architecture, distributed satellite computing network architecture, and reconfigurable satellite formation flying, to enable flexible and scalable communication, computing and control. The DSIN faces challenges from network heterogeneity, unpredictable channel dynamics, sparse resources, and decentralized collaboration frameworks. To address these issues, a series of enabling technologies is identified, including channel modeling and estimation, cloud-native distributed MIMO cooperation, grant-free massive access, network routing, and the proper combination of all these diversity techniques. Furthermore, to heighten the overall resource efficiency, the cross-layer optimization techniques are further developed to meet upper-layer deterministic, adaptive and secure information services requirements. In addition, emerging research directions and new opportunities are highlighted on the way to achieving the DSIN vision.

Zhenghao Lin, Zihao Tang, Xiao Liu et al.

We introduce Sigma, an efficient large language model specialized for the system domain, empowered by a novel architecture including DiffQKV attention, and pre-trained on our meticulously collected system domain data. DiffQKV attention significantly enhances the inference efficiency of Sigma by optimizing the Query (Q), Key (K), and Value (V) components in the attention mechanism differentially, based on their varying impacts on the model performance and efficiency indicators. Specifically, we (1) conduct extensive experiments that demonstrate the model's varying sensitivity to the compression of K and V components, leading to the development of differentially compressed KV, and (2) propose augmented Q to expand the Q head dimension, which enhances the model's representation capacity with minimal impacts on the inference speed. Rigorous theoretical and empirical analyses reveal that DiffQKV attention significantly enhances efficiency, achieving up to a 33.36% improvement in inference speed over the conventional grouped-query attention (GQA) in long-context scenarios. We pre-train Sigma on 6T tokens from various sources, including 19.5B system domain data that we carefully collect and 1T tokens of synthesized and rewritten data. In general domains, Sigma achieves comparable performance to other state-of-arts models. In the system domain, we introduce the first comprehensive benchmark AIMicius, where Sigma demonstrates remarkable performance across all tasks, significantly outperforming GPT-4 with an absolute improvement up to 52.5%.

Yang Li, Mingxuan Luo, Yeyun Gong et al.

Supervised fine-tuning with synthesized instructions has been a common practice for adapting LLMs to domain-specific QA tasks. However, the synthesized instructions deviate from real user questions and expected answers. This study proposes a novel framework called DeepThink to generate high-quality instructions. DeepThink first generates a few seed questions to mimic actual user questions, simulates conversations to uncover the hidden user needs, and refines the answer by conversational contexts and the retrieved documents for more comprehensive answers. Experiments demonstrate that DeepThink achieves an average performance improvement of 7.92% compared to a GPT-4-turbo+RAG-based assistant on the real user test set in the advertising domain across dimensions such as relevance, completeness, clarity, accuracy, and actionability.

Jian Jiao, Yu Dai, Hefei Mei et al.

Recent video class-incremental learning usually excessively pursues the accuracy of the newly seen classes and relies on memory sets to mitigate catastrophic forgetting of the old classes. However, limited storage only allows storing a few representative videos. So we propose SNRO, which slightly shifts the features of new classes to remember old classes. Specifically, SNRO contains Examples Sparse(ES) and Early Break(EB). ES decimates at a lower sample rate to build memory sets and uses interpolation to align those sparse frames in the future. By this, SNRO stores more examples under the same memory consumption and forces the model to focus on low-semantic features which are harder to be forgotten. EB terminates the training at a small epoch, preventing the model from overstretching into the high-semantic space of the current task. Experiments on UCF101, HMDB51, and UESTC-MMEA-CL datasets show that SNRO performs better than other approaches while consuming the same memory consumption.

Yaoxiang Wang, Qingguo Hu, Yucheng Ding et al.

Mixture-of-Experts (MoE) has emerged as a promising paradigm for efficiently scaling large language models without a proportional increase in computational cost. However, the standard training strategy of Top-K router prevents MoE models from realizing their full potential for elastic inference. When the number of activated experts is altered at inference time, these models exhibit precipitous performance degradation. In this work, we introduce Matryoshka MoE (M-MoE), a training framework that instills a coarse-to-fine structure directly into the expert ensemble. By systematically varying the number of activated experts during training, M-MoE compels the model to learn a meaningful ranking: top-ranked experts collaborate to provide essential, coarse-grained capabilities, while subsequent experts add progressively finer-grained detail. We explore this principle at multiple granularities, identifying a layer-wise randomization strategy as the most effective. Our experiments demonstrate that a single M-MoE model achieves remarkable elasticity, with its performance at various expert counts closely matching that of an entire suite of specialist models, but at only a fraction of the total training cost. This flexibility not only unlocks elastic inference but also enables optimizing performance by allocating different computational budgets to different model layers. Our work paves the way for more practical and adaptable deployments of large-scale MoE models.

Shiyao Jiang, Jian Jiao, Xingjian Zhang et al.

With the emergence of diverse and massive data in the upcoming sixth-generation (6G) networks, the task-agnostic semantic communication system is regarded to provide robust intelligent services. In this paper, we propose a task-agnostic learnable weighted-knowledge base semantic communication (TALSC) framework for robust image transmission to address the real-world heterogeneous data bias in KB, including label flipping noise and class imbalance. The TALSC framework incorporates a sample confidence module (SCM) as meta-learner and the semantic coding networks as learners. The learners are updated based on the empirical knowledge provided by the learnable weighted-KB (LW-KB). Meanwhile, the meta-learner evaluates the significance of samples according to the task loss feedback, and adjusts the update strategy of learners to enhance the robustness in semantic recovery for unknown tasks. To strike a balance between SCM parameters and precision of significance evaluation, we design an SCM-grid extension (SCM-GE) approach by embedding the Kolmogorov-Arnold networks (KAN) within SCM, which leverages the concept of spline refinement in KAN and enables scalable SCM with customizable granularity without retraining. Simulations demonstrate that the TALSC framework effectively mitigates the effects of flipping noise and class imbalance in task-agnostic image semantic communication, achieving at least 12% higher semantic recovery accuracy (SRA) and multi-scale structural similarity (MS-SSIM) compared to state-of-the-art methods.

Xiao Liang, Xinyu Hu, Simiao Zuo et al.

Large Language Models (LLMs) have shown superior performance in various applications and fields. To achieve better performance on specialized domains such as law and advertisement, LLMs are often continue pre-trained on in-domain data. However, existing approaches suffer from two major issues. First, in-domain data are scarce compared with general domain-agnostic data. Second, data used for continual pre-training are not task-aware, such that they may not be helpful to downstream applications. We propose TRAIT, a task-oriented in-domain data augmentation framework. Our framework is divided into two parts: in-domain data selection and task-oriented synthetic passage generation. The data selection strategy identifies and selects a large amount of in-domain data from general corpora, and thus significantly enriches domain knowledge in the continual pre-training data. The synthetic passages contain guidance on how to use domain knowledge to answer questions about downstream tasks. By training on such passages, the model aligns with the need of downstream applications. We adapt LLMs to two domains: advertisement and math. On average, TRAIT improves LLM performance by 8% in the advertisement domain and 7.5% in the math domain.

Amit Sharma, Hua Li, Xue Li et al.

Given an input query, a recommendation model is trained using user feedback data (e.g., click data) to output a ranked list of items. In real-world systems, besides accuracy, an important consideration for a new model is novelty of its top-k recommendations w.r.t. an existing deployed model. However, novelty of top-k items is a difficult goal to optimize a model for, since it involves a non-differentiable sorting operation on the model's predictions. Moreover, novel items, by definition, do not have any user feedback data. Given the semantic capabilities of large language models, we address these problems using a reinforcement learning (RL) formulation where large language models provide feedback for the novel items. However, given millions of candidate items, the sample complexity of a standard RL algorithm can be prohibitively high. To reduce sample complexity, we reduce the top-k list reward to a set of item-wise rewards and reformulate the state space to consist of <query, item> tuples such that the action space is reduced to a binary decision; and show that this reformulation results in a significantly lower complexity when the number of items is large. We evaluate the proposed algorithm on improving novelty for a query-ad recommendation task on a large-scale search engine. Compared to supervised finetuning on recent <query, ad> pairs, the proposed RL-based algorithm leads to significant novelty gains with minimal loss in recall. We obtain similar results on the ORCAS query-webpage matching dataset and a product recommendation dataset based on Amazon reviews.

Gurusha Juneja, Gautam Jajoo, Nagarajan Natarajan et al.

Given a task in the form of a basic description and its training examples, prompt optimization is the problem of synthesizing the given information into a text prompt for a large language model. Humans solve this problem by also considering the different facets that define a task (e.g., counter-examples, explanations, analogies) and including them in the prompt. However, it is unclear whether existing algorithmic approaches, based on iteratively editing a given prompt or automatically selecting a few in-context examples, can cover the multiple facets required to solve a complex task. In this work, we view prompt optimization as that of learning multiple facets of a task from a set of training examples. We exploit structure in the prompt optimization problem and break down a prompt into loosely coupled semantic sections. The proposed algorithm, UniPrompt, (1) clusters the input space and uses clustered batches so that each batch likely corresponds to a different facet of the task, and (2) utilizes a feedback mechanism to propose adding, editing or deleting a section, which in turn is aggregated over a batch to capture generalizable facets. Empirical evaluation on multiple datasets and a real-world task shows that prompts generated using \shortname{} obtain higher accuracy than human-tuned prompts and those from state-of-the-art methods. In particular, our algorithm can generate long, complex prompts that existing methods are unable to generate. Code for UniPrompt is available at https://aka.ms/uniprompt.

Ravisri Valluri, Akash Kumar Mohankumar, Kushal Dave et al.

Generative Retrieval introduces a new approach to Information Retrieval by reframing it as a constrained generation task, leveraging recent advancements in Autoregressive (AR) language models. However, AR-based Generative Retrieval methods suffer from high inference latency and cost compared to traditional dense retrieval techniques, limiting their practical applicability. This paper investigates fully Non-autoregressive (NAR) language models as a more efficient alternative for generative retrieval. While standard NAR models alleviate latency and cost concerns, they exhibit a significant drop in retrieval performance (compared to AR models) due to their inability to capture dependencies between target tokens. To address this, we question the conventional choice of limiting the target token space to solely words or sub-words. We propose PIXAR, a novel approach that expands the target vocabulary of NAR models to include multi-word entities and common phrases (up to 5 million tokens), thereby reducing token dependencies. PIXAR employs inference optimization strategies to maintain low inference latency despite the significantly larger vocabulary. Our results demonstrate that PIXAR achieves a relative improvement of 31.0% in MRR@10 on MS MARCO and 23.2% in Hits@5 on Natural Questions compared to standard NAR models with similar latency and cost. Furthermore, online A/B experiments on a large commercial search engine show that PIXAR increases ad clicks by 5.08% and revenue by 4.02%.

Ting Jiang, Jian Jiao, Shaohan Huang et al.

We propose PromptBERT, a novel contrastive learning method for learning better sentence representation. We firstly analyze the drawback of current sentence embedding from original BERT and find that it is mainly due to the static token embedding bias and ineffective BERT layers. Then we propose the first prompt-based sentence embeddings method and discuss two prompt representing methods and three prompt searching methods to make BERT achieve better sentence embeddings. Moreover, we propose a novel unsupervised training objective by the technology of template denoising, which substantially shortens the performance gap between the supervised and unsupervised settings. Extensive experiments show the effectiveness of our method. Compared to SimCSE, PromptBert achieves 2.29 and 2.58 points of improvement based on BERT and RoBERTa in the unsupervised setting.

Haonan Li, Yeyun Gong, Jian Jiao et al.

Pre-trained language models have led to substantial gains over a broad range of natural language processing (NLP) tasks, but have been shown to have limitations for natural language generation tasks with high-quality requirements on the output, such as commonsense generation and ad keyword generation. In this work, we present a novel Knowledge Filtering and Contrastive learning Network (KFCNet) which references external knowledge and achieves better generation performance. Specifically, we propose a BERT-based filter model to remove low-quality candidates, and apply contrastive learning separately to each of the encoder and decoder, within a general encoder--decoder architecture. The encoder contrastive module helps to capture global target semantics during encoding, and the decoder contrastive module enhances the utility of retrieved prototypes while learning general features. Extensive experiments on the CommonGen benchmark show that our model outperforms the previous state of the art by a large margin: +6.6 points (42.5 vs. 35.9) for BLEU-4, +3.7 points (33.3 vs. 29.6) for SPICE, and +1.3 points (18.3 vs. 17.0) for CIDEr. We further verify the effectiveness of the proposed contrastive module on ad keyword generation, and show that our model has potential commercial value.

Zhihao Fan, Yeyun Gong, Dayiheng Liu et al.

Transformer is an attention-based neural network, which consists of two sublayers, namely, Self-Attention Network (SAN) and Feed-Forward Network (FFN). Existing research explores to enhance the two sublayers separately to improve the capability of Transformer for text representation. In this paper, we present a novel understanding of SAN and FFN as Mask Attention Networks (MANs) and show that they are two special cases of MANs with static mask matrices. However, their static mask matrices limit the capability for localness modeling in text representation learning. We therefore introduce a new layer named dynamic mask attention network (DMAN) with a learnable mask matrix which is able to model localness adaptively. To incorporate advantages of DMAN, SAN, and FFN, we propose a sequential layered structure to combine the three types of layers. Extensive experiments on various tasks, including neural machine translation and text summarization demonstrate that our model outperforms the original Transformer.

Weizhen Qi, Yeyun Gong, Jian Jiao et al.

In this paper, we propose BANG, a new pretraining model to Bridge the gap between Autoregressive (AR) and Non-autoregressive (NAR) Generation. AR and NAR generation can be uniformly regarded as to what extent previous tokens can be attended, and BANG bridges AR and NAR generation by designing a novel model structure for large-scale pretraining. The pretrained BANG model can simultaneously support AR, NAR and semi-NAR generation to meet different requirements. Experiments on question generation (SQuAD 1.1), summarization (XSum) and dialogue generation (PersonaChat) show that BANG improves NAR and semi-NAR performance significantly as well as attaining comparable performance with strong AR pretrained models. Compared with the semi-NAR strong baselines, BANG achieves absolute improvements of 14.01 and 5.24 in the overall scores of SQuAD 1.1 and XSum, respectively. In addition, BANG achieves absolute improvements of 10.73, 6.39 and 5.90 in the overall scores of SQuAD, XSUM and PersonaChat respectively compared with the strong NAR baselines.

Zhihao Fan, Yeyun Gong, Zhongyu Wei et al.

Commonsense generation aims at generating plausible everyday scenario description based on a set of provided concepts. Digging the relationship of concepts from scratch is non-trivial, therefore, we retrieve prototypes from external knowledge to assist the understanding of the scenario for better description generation. We integrate two additional modules, namely position indicator and scaling module, into the pretrained encoder-decoder model for prototype modeling to enhance the knowledge injection procedure. We conduct experiment on CommonGen benchmark, and experimental results show that our method significantly improves the performance on all the metrics.

Weizhen Qi, Yeyun Gong, Yu Yan et al.

In a sponsored search engine, generative retrieval models are recently proposed to mine relevant advertisement keywords for users' input queries. Generative retrieval models generate outputs token by token on a path of the target library prefix tree (Trie), which guarantees all of the generated outputs are legal and covered by the target library. In actual use, we found several typical problems caused by Trie-constrained searching length. In this paper, we analyze these problems and propose a looking ahead strategy for generative retrieval models named ProphetNet-Ads. ProphetNet-Ads improves the retrieval ability by directly optimizing the Trie-constrained searching space. We build a dataset from a real-word sponsored search engine and carry out experiments to analyze different generative retrieval models. Compared with Trie-based LSTM generative retrieval model proposed recently, our single model result and integrated result improve the recall by 15.58\% and 18.8\% respectively with beam size 5. Case studies further demonstrate how these problems are alleviated by ProphetNet-Ads clearly.

Sanxing Chen, Xiaodong Liu, Jianfeng Gao et al.

This paper examines the challenging problem of learning representations of entities and relations in a complex multi-relational knowledge graph. We propose HittER, a Hierarchical Transformer model to jointly learn Entity-relation composition and Relational contextualization based on a source entity's neighborhood. Our proposed model consists of two different Transformer blocks: the bottom block extracts features of each entity-relation pair in the local neighborhood of the source entity and the top block aggregates the relational information from outputs of the bottom block. We further design a masked entity prediction task to balance information from the relational context and the source entity itself. Experimental results show that HittER achieves new state-of-the-art results on multiple link prediction datasets. We additionally propose a simple approach to integrate HittER into BERT and demonstrate its effectiveness on two Freebase factoid question answering datasets.

Wenhao Lu, Jian Jiao, Ruofei Zhang

Pre-trained language models like BERT have achieved great success in a wide variety of NLP tasks, while the superior performance comes with high demand in computational resources, which hinders the application in low-latency IR systems. We present TwinBERT model for effective and efficient retrieval, which has twin-structured BERT-like encoders to represent query and document respectively and a crossing layer to combine the embeddings and produce a similarity score. Different from BERT, where the two input sentences are concatenated and encoded together, TwinBERT decouples them during encoding and produces the embeddings for query and document independently, which allows document embeddings to be pre-computed offline and cached in memory. Thereupon, the computation left for run-time is from the query encoding and query-document crossing only. This single change can save large amount of computation time and resources, and therefore significantly improve serving efficiency. Moreover, a few well-designed network layers and training strategies are proposed to further reduce computational cost while at the same time keep the performance as remarkable as BERT model. Lastly, we develop two versions of TwinBERT for retrieval and relevance tasks correspondingly, and both of them achieve close or on-par performance to BERT-Base model. The model was trained following the teacher-student framework and evaluated with data from one of the major search engines. Experimental results showed that the inference time was significantly reduced and was firstly controlled around 20ms on CPUs while at the same time the performance gain from fine-tuned BERT-Base model was mostly retained. Integration of the models into production systems also demonstrated remarkable improvements on relevance metrics with negligible influence on latency.

Jian Jiao, Jichao Jiao, Yaokai Mo et al.

This paper proposes a new framework to solve the problem of monocular visual odometry, called MagicVO . Based on Convolutional Neural Network (CNN) and Bi-directional LSTM (Bi-LSTM), MagicVO outputs a 6-DoF absolute-scale pose at each position of the camera with a sequence of continuous monocular images as input. It not only utilizes the outstanding performance of CNN in image feature processing to extract the rich features of image frames fully but also learns the geometric relationship from image sequences pre and post through Bi-LSTM to get a more accurate prediction. A pipeline of the MagicVO is shown in Fig. 1. The MagicVO system is end-to-end, and the results of experiments on the KITTI dataset and the ETH-asl cla dataset show that MagicVO has a better performance than traditional visual odometry (VO) systems in the accuracy of pose and the generalization ability.

Ying Shan, Jian Jiao, Jie Zhu et al.

Rapid advances in GPU hardware and multiple areas of Deep Learning open up a new opportunity for billion-scale information retrieval with exhaustive search. Building on top of the powerful concept of semantic learning, this paper proposes a Recurrent Binary Embedding (RBE) model that learns compact representations for real-time retrieval. The model has the unique ability to refine a base binary vector by progressively adding binary residual vectors to meet the desired accuracy. The refined vector enables efficient implementation of exhaustive similarity computation with bit-wise operations, followed by a near- lossless k-NN selection algorithm, also proposed in this paper. The proposed algorithms are integrated into an end-to-end multi-GPU system that retrieves thousands of top items from over a billion candidates in real-time. The RBE model and the retrieval system were evaluated with data from a major paid search engine. When measured against the state-of-the-art model for binary representation and the full precision model for semantic embedding, RBE significantly outperformed the former, and filled in over 80% of the AUC gap in-between. Experiments comparing with our production retrieval system also demonstrated superior performance. While the primary focus of this paper is to build RBE based on a particular class of semantic models, generalizing to other types is straightforward, as exemplified by two different models at the end of the paper.