Zhefeng Wang

Qi Liu, Mingdi Sun, Yongyi He et al.

Supervised fine-tuning (SFT) followed by reinforcement learning (RL) has become a standard post-training paradigm for large language models. This paradigm provides a cold-start for RL exploration, avoiding the inefficiency of pure RL where on-policy sampling yields insufficient positive samples. However, in practice, existing approaches often use a small amount of data for SFT initialization compared to the RL phase, which can cause the model to fit the limited samples and shift away from its pre-trained distribution. This distribution shift impedes the model's ability to effectively explore during subsequent RL training. To address this challenge, we propose that in low-data regimes, SFT should prioritize activating task-relevant capabilities rather than memorizing specific content. Along this line, we propose EKSFT (Entropy-KL Selective Fine-Tuning), which selectively masks tokens that exhibit either high entropy or high KL divergence from a reference model. By excluding these high-uncertainty, distribution-shifting tokens from imitation, EKSFT injects task-specific knowledge while preserving the integrity of the model's pre-trained distribution. Empirical evaluations on mathematical reasoning benchmarks demonstrate that EKSFT consistently outperforms standard SFT. Further RL fine-tuning from the EKSFT model yields consistently better post-RL performance, indicating improved exploration for the RL stage. Our codes and datasets are available at https://github.com/MINE-USTC/EKSFT.

Tong Zhu, Junfei Ren, Zijian Yu et al.

Sharing knowledge between information extraction tasks has always been a challenge due to the diverse data formats and task variations. Meanwhile, this divergence leads to information waste and increases difficulties in building complex applications in real scenarios. Recent studies often formulate IE tasks as a triplet extraction problem. However, such a paradigm does not support multi-span and n-ary extraction, leading to weak versatility. To this end, we reorganize IE problems into unified multi-slot tuples and propose a universal framework for various IE tasks, namely Mirror. Specifically, we recast existing IE tasks as a multi-span cyclic graph extraction problem and devise a non-autoregressive graph decoding algorithm to extract all spans in a single step. It is worth noting that this graph structure is incredibly versatile, and it supports not only complex IE tasks, but also machine reading comprehension and classification tasks. We manually construct a corpus containing 57 datasets for model pretraining, and conduct experiments on 30 datasets across 8 downstream tasks. The experimental results demonstrate that our model has decent compatibility and outperforms or reaches competitive performance with SOTA systems under few-shot and zero-shot settings. The code, model weights, and pretraining corpus are available at https://github.com/Spico197/Mirror .

Tong Zhu, Guoliang Zhang, Zechang Li et al.

Sentence-by-sentence information extraction from long documents is an exhausting and error-prone task. As the indicator of document skeleton, catalogs naturally chunk documents into segments and provide informative cascade semantics, which can help to reduce the search space. Despite their usefulness, catalogs are hard to be extracted without the assist from external knowledge. For documents that adhere to a specific template, regular expressions are practical to extract catalogs. However, handcrafted heuristics are not applicable when processing documents from different sources with diverse formats. To address this problem, we build a large manually annotated corpus, which is the first dataset for the Catalog Extraction from Documents (CED) task. Based on this corpus, we propose a transition-based framework for parsing documents into catalog trees. The experimental results demonstrate that our proposed method outperforms baseline systems and shows a good ability to transfer. We believe the CED task could fill the gap between raw text segments and information extraction tasks on extremely long documents. Data and code are available at \url{https://github.com/Spico197/CatalogExtraction}

Likang Wu, Zhi Li, Hongke Zhao et al.

Zero-Shot Learning (ZSL), which aims at automatically recognizing unseen objects, is a promising learning paradigm to understand new real-world knowledge for machines continuously. Recently, the Knowledge Graph (KG) has been proven as an effective scheme for handling the zero-shot task with large-scale and non-attribute data. Prior studies always embed relationships of seen and unseen objects into visual information from existing knowledge graphs to promote the cognitive ability of the unseen data. Actually, real-world knowledge is naturally formed by multimodal facts. Compared with ordinary structural knowledge from a graph perspective, multimodal KG can provide cognitive systems with fine-grained knowledge. For example, the text description and visual content can depict more critical details of a fact than only depending on knowledge triplets. Unfortunately, this multimodal fine-grained knowledge is largely unexploited due to the bottleneck of feature alignment between different modalities. To that end, we propose a multimodal intensive ZSL framework that matches regions of images with corresponding semantic embeddings via a designed dense attention module and self-calibration loss. It makes the semantic transfer process of our ZSL framework learns more differentiated knowledge between entities. Our model also gets rid of the performance limitation of only using rough global features. We conduct extensive experiments and evaluate our model on large-scale real-world data. The experimental results clearly demonstrate the effectiveness of the proposed model in standard zero-shot classification tasks.

Yingjie Gu, Xiaoye Qu, Zhefeng Wang et al.

Recent years have witnessed the improving performance of Chinese Named Entity Recognition (NER) from proposing new frameworks or incorporating word lexicons. However, the inner composition of entity mentions in character-level Chinese NER has been rarely studied. Actually, most mentions of regular types have strong name regularity. For example, entities end with indicator words such as "company" or "bank" usually belong to organization. In this paper, we propose a simple but effective method for investigating the regularity of entity spans in Chinese NER, dubbed as Regularity-Inspired reCOgnition Network (RICON). Specifically, the proposed model consists of two branches: a regularity-aware module and a regularityagnostic module. The regularity-aware module captures the internal regularity of each span for better entity type prediction, while the regularity-agnostic module is employed to locate the boundary of entities and relieve the excessive attention to span regularity. An orthogonality space is further constructed to encourage two modules to extract different aspects of regularity features. To verify the effectiveness of our method, we conduct extensive experiments on three benchmark datasets and a practical medical dataset. The experimental results show that our RICON significantly outperforms previous state-of-the-art methods, including various lexicon-based methods.

Xiaoye Qu, Yingjie Gu, Qingrong Xia et al.

As more and more Arabic texts emerged on the Internet, extracting important information from these Arabic texts is especially useful. As a fundamental technology, Named entity recognition (NER) serves as the core component in information extraction technology, while also playing a critical role in many other Natural Language Processing (NLP) systems, such as question answering and knowledge graph building. In this paper, we provide a comprehensive review of the development of Arabic NER, especially the recent advances in deep learning and pre-trained language model. Specifically, we first introduce the background of Arabic NER, including the characteristics of Arabic and existing resources for Arabic NER. Then, we systematically review the development of Arabic NER methods. Traditional Arabic NER systems focus on feature engineering and designing domain-specific rules. In recent years, deep learning methods achieve significant progress by representing texts via continuous vector representations. With the growth of pre-trained language model, Arabic NER yields better performance. Finally, we conclude the method gap between Arabic NER and NER methods from other languages, which helps outline future directions for Arabic NER.

Xiaoye Qu, Jun Zeng, Daizong Liu et al.

Distantly-Supervised Named Entity Recognition (DS-NER) effectively alleviates the data scarcity problem in NER by automatically generating training samples. Unfortunately, the distant supervision may induce noisy labels, thus undermining the robustness of the learned models and restricting the practical application. To relieve this problem, recent works adopt self-training teacher-student frameworks to gradually refine the training labels and improve the generalization ability of NER models. However, we argue that the performance of the current self-training frameworks for DS-NER is severely underestimated by their plain designs, including both inadequate student learning and coarse-grained teacher updating. Therefore, in this paper, we make the first attempt to alleviate these issues by proposing: (1) adaptive teacher learning comprised of joint training of two teacher-student networks and considering both consistent and inconsistent predictions between two teachers, thus promoting comprehensive student learning. (2) fine-grained student ensemble that updates each fragment of the teacher model with a temporal moving average of the corresponding fragment of the student, which enhances consistent predictions on each model fragment against noise. To verify the effectiveness of our proposed method, we conduct experiments on four DS-NER datasets. The experimental results demonstrate that our method significantly surpasses previous SOTA methods.

Abbas Ghaddar, Yimeng Wu, Sunyam Bagga et al.

There is a growing body of work in recent years to develop pre-trained language models (PLMs) for the Arabic language. This work concerns addressing two major problems in existing Arabic PLMs which constraint progress of the Arabic NLU and NLG fields.First, existing Arabic PLMs are not well-explored and their pre-trainig can be improved significantly using a more methodical approach. Second, there is a lack of systematic and reproducible evaluation of these models in the literature. In this work, we revisit both the pre-training and evaluation of Arabic PLMs. In terms of pre-training, we explore improving Arabic LMs from three perspectives: quality of the pre-training data, size of the model, and incorporating character-level information. As a result, we release three new Arabic BERT-style models ( JABER, Char-JABER, and SABER), and two T5-style models (AT5S and AT5B). In terms of evaluation, we conduct a comprehensive empirical study to systematically evaluate the performance of existing state-of-the-art models on ALUE that is a leaderboard-powered benchmark for Arabic NLU tasks, and on a subset of the ARGEN benchmark for Arabic NLG tasks. We show that our models significantly outperform existing Arabic PLMs and achieve a new state-of-the-art performance on discriminative and generative Arabic NLU and NLG tasks. Our models and source code to reproduce of results will be made available shortly.

Asaad Alghamdi, Xinyu Duan, Wei Jiang et al.

Developing monolingual large Pre-trained Language Models (PLMs) is shown to be very successful in handling different tasks in Natural Language Processing (NLP). In this work, we present AraMUS, the largest Arabic PLM with 11B parameters trained on 529GB of high-quality Arabic textual data. AraMUS achieves state-of-the-art performances on a diverse set of Arabic classification and generative tasks. Moreover, AraMUS shows impressive few-shot learning abilities compared with the best existing Arabic PLMs.

Mingqi Gao, Xiaojun Wan, Jia Su et al.

Factuality is important to dialogue summarization. Factual error correction (FEC) of model-generated summaries is one way to improve factuality. Current FEC evaluation that relies on factuality metrics is not reliable and detailed enough. To address this problem, we are the first to manually annotate a FEC dataset for dialogue summarization containing 4000 items and propose FERRANTI, a fine-grained evaluation framework based on reference correction that automatically evaluates the performance of FEC models on different error categories. Using this evaluation framework, we conduct sufficient experiments with FEC approaches under a variety of settings and find the best training modes and significant differences in the performance of the existing approaches on different factual error categories.

Lei Zhang, Zhenghua Li, Shilin Zhou et al.

Inspired by early research on exploring naturally annotated data for Chinese word segmentation (CWS), and also by recent research on integration of speech and text processing, this work for the first time proposes to mine word boundaries from parallel speech/text data. First we collect parallel speech/text data from two Internet sources that are related with CWS data used in our experiments. Then, we obtain character-level alignments and design simple heuristic rules for determining word boundaries according to pause duration between adjacent characters. Finally, we present an effective complete-then-train strategy that can better utilize extra naturally annotated data for model training. Experiments demonstrate our approach can significantly boost CWS performance in both cross-domain and low-resource scenarios.

Houquan Zhou, Yang Hou, Zhenghua Li et al.

While recent advancements in large language models (LLMs) bring us closer to achieving artificial general intelligence, the question persists: Do LLMs truly understand language, or do they merely mimic comprehension through pattern recognition? This study seeks to explore this question through the lens of syntax, a crucial component of sentence comprehension. Adopting a natural language question-answering (Q&A) scheme, we craft questions targeting nine syntactic knowledge points that are most closely related to sentence comprehension. Experiments conducted on 24 LLMs suggest that most have a limited grasp of syntactic knowledge, exhibiting notable discrepancies across different syntactic knowledge points. In particular, questions involving prepositional phrase attachment pose the greatest challenge, whereas those concerning adjectival modifier and indirect object are relatively easier for LLMs to handle. Furthermore, a case study on the training dynamics of the LLMs reveals that the majority of syntactic knowledge is learned during the initial stages of training, hinting that simply increasing the number of training tokens may not be the `silver bullet' for improving the comprehension ability of LLMs.

Yanggan Gu, Yang Hou, Zhefeng Wang et al.

This work revisits the topic of jointly parsing constituency and dependency trees, i.e., to produce compatible constituency and dependency trees simultaneously for input sentences, which is attractive considering that the two types of trees are complementary in representing syntax. The original work of Zhou and Zhao (2019) performs joint parsing only at the inference phase. They train two separate parsers under the multi-task learning framework (i.e., one shared encoder and two independent decoders). They design an ad-hoc dynamic programming-based decoding algorithm of $O(n^5)$ time complexity for finding optimal compatible tree pairs. Compared to their work, we make progress in three aspects: (1) adopting a much more efficient decoding algorithm of $O(n^4)$ time complexity, (2) exploring joint modeling at the training phase, instead of only at the inference phase, (3) proposing high-order scoring components to promote constituent-dependency interaction. We conduct experiments and analysis on seven languages, covering both rich-resource and low-resource scenarios. Results and analysis show that joint modeling leads to a modest overall performance boost over separate modeling, but substantially improves the complete matching ratio of whole trees, thanks to the explicit modeling of tree compatibility.

Yixin Ji, Yang Xiang, Juntao Li et al.

As large language models (LLMs) are widely applied across various fields, model compression has become increasingly crucial for reducing costs and improving inference efficiency. Post-training pruning is a promising method that does not require resource-intensive iterative training and only needs a small amount of calibration data to assess the importance of parameters. Recent research has enhanced post-training pruning from different aspects but few of them systematically explore the effects of calibration data, and it is unclear if there exist better calibration data construction strategies. We fill this blank and surprisingly observe that calibration data is also crucial to post-training pruning, especially for high sparsity. Through controlled experiments on important influence factors of calibration data, including the pruning settings, the amount of data, and its similarity with pre-training data, we observe that a small size of data is adequate, and more similar data to its pre-training stage can yield better performance. As pre-training data is usually inaccessible for advanced LLMs, we further provide a self-generating calibration data synthesis strategy to construct feasible calibration data. Experimental results on recent strong open-source LLMs (e.g., DCLM, and LLaMA-3) show that the proposed strategy can enhance the performance of strong pruning methods (e.g., Wanda, DSnoT, OWL) by a large margin (up to $2.68\%$). Code is available at https://github.com/Dereck0602/calibration_data.

Zhongyang Guo, Guanran Jiang, Zhongdan Zhang et al.

This paper introduces "Shai" a 10B level large language model specifically designed for the asset management industry, built upon an open-source foundational model. With continuous pre-training and fine-tuning using a targeted corpus, Shai demonstrates enhanced performance in tasks relevant to its domain, outperforming baseline models. Our research includes the development of an innovative evaluation framework, which integrates professional qualification exams, tailored tasks, open-ended question answering, and safety assessments, to comprehensively assess Shai's capabilities. Furthermore, we discuss the challenges and implications of utilizing large language models like GPT-4 for performance assessment in asset management, suggesting a combination of automated evaluation and human judgment. Shai's development, showcasing the potential and versatility of 10B-level large language models in the financial sector with significant performance and modest computational requirements, hopes to provide practical insights and methodologies to assist industry peers in their similar endeavors.

Gursimran Singh, Xinglu Wang, Yifan Hu et al.

Large Multimodal Models (LMMs) extend Large Language Models (LLMs) by handling diverse inputs such as images, audio, and video, but at the cost of adding a multimodal encoding stage that increases both computational and memory overhead. This step negatively affects key Service Level Objectives (SLOs), such as time to first token (TTFT) and time per output token (TPOT). We introduce Encode-Prefill-Decode (EPD) Disaggregation, a novel framework that separates the encoding, prefill, and decode stages onto dedicated resources. Unlike current systems, which bundle encoding and prefill together, our approach decouples these steps, unlocking new opportunities and optimizations. These include a mechanism to cache multimedia tokens for efficient transfer, a novel way to parallelize the encoding load within a request, a module for optimal resource allocation for disaggregated serving, and a novel role-switching method to handle changing workload characteristics. Experimental evaluations with popular LMMs show substantial gains in memory efficiency (up to 15x lower peak memory utilization), batch sizes (up to 22x larger), 10x more images per request, and 2.2x larger KV caches. Furthermore, it leads to significant improvements in SLO attainment (up to 90-100% improvement) and TTFT (up to 71% reduction), compared to systems that do not disaggregate. The code is available at https://github.com/vbdi/epdserve.

Pengxiang Zhao, Hanyu Hu, Ping Li et al.

Pruning is a critical strategy for compressing trained large language models (LLMs), aiming at substantial memory conservation and computational acceleration without compromising performance. However, existing pruning methods often necessitate inefficient retraining for billion-scale LLMs or rely on heuristic methods such as the optimal brain surgeon framework, which degrade performance. In this paper, we introduce FISTAPruner, the first post-training pruner based on convex optimization models and algorithms. Specifically, we propose a convex optimization model incorporating $\ell_1$ norm to induce sparsity and utilize the FISTA solver for optimization. FISTAPruner incorporates an intra-layer cumulative error correction mechanism and supports parallel pruning. We comprehensively evaluate FISTAPruner on models such as OPT, LLaMA, LLaMA-2, and LLaMA-3 with 125M to 70B parameters under unstructured and 2:4 semi-structured sparsity, demonstrating superior performance over existing state-of-the-art methods across various language benchmarks.

Jikai Wang, Yi Su, Juntao Li et al.

Autoregressive language models demonstrate excellent performance in various scenarios. However, the inference efficiency is limited by its one-step-one-word generation mode, which has become a pressing problem recently as the models become increasingly larger. Speculative decoding employs a "draft and then verify" mechanism to allow multiple tokens to be generated in one step, realizing lossless acceleration. Existing methods mainly adopt fixed heuristic draft structures, which fail to adapt to different situations to maximize the acceptance length during verification. To alleviate this dilemma, we proposed OPT-Tree, an algorithm to construct adaptive and scalable draft trees. It searches the optimal tree structure that maximizes the mathematical expectation of the acceptance length in each decoding step. Experimental results reveal that OPT-Tree outperforms the existing draft structures and achieves a speed-up ratio of up to 3.2 compared with autoregressive decoding. If the draft model is powerful enough and the node budget is sufficient, it can generate more than ten tokens in a single step. Our code is available at https://github.com/Jikai0Wang/OPT-Tree.

Tong Zhu, Xiaoye Qu, Wenliang Chen et al.

Most previous studies of document-level event extraction mainly focus on building argument chains in an autoregressive way, which achieves a certain success but is inefficient in both training and inference. In contrast to the previous studies, we propose a fast and lightweight model named as PTPCG. In our model, we design a novel strategy for event argument combination together with a non-autoregressive decoding algorithm via pruned complete graphs, which are constructed under the guidance of the automatically selected pseudo triggers. Compared to the previous systems, our system achieves competitive results with 19.8\% of parameters and much lower resource consumption, taking only 3.8\% GPU hours for training and up to 8.5 times faster for inference. Besides, our model shows superior compatibility for the datasets with (or without) triggers and the pseudo triggers can be the supplements for annotated triggers to make further improvements. Codes are available at https://github.com/Spico197/DocEE .

Yanke Yu, Jin Li, Ying Sun et al.

Understanding the internal functional organization of Large Language Models (LLMs) is crucial for improving their trustworthiness and performance. However, how LLMs organize different functions into modules remains highly unexplored. To bridge this gap, we formulate a functional module discovery problem and propose an Unsupervised LLM Cross-layer MOdule Discovery (ULCMOD) framework that simultaneously disentangles the large set of neurons in the entire LLM into modules while discovering the topics of input samples related to these modules. Our framework introduces a novel objective function and an efficient Iterative Decoupling (IterD) algorithm. Extensive experiments show that our method discovers high-quality, disentangled modules that capture more meaningful semantic information and achieve superior performance in various downstream tasks. Moreover, our qualitative analysis reveals that the discovered modules show semantic coherence, correspond to interpretable specializations, and a clear spatial and hierarchical organization within the LLM. Our work provides a novel tool for interpreting the functional modules of LLMs, filling a critical blank in LLM's interpretability research.

Yixin Ji, Yang Xiang, Juntao Li et al.

In recent years, large language models (LLMs) have driven advances in natural language processing. Still, their growing scale has increased the computational burden, necessitating a balance between efficiency and performance. Low-rank compression, a promising technique, reduces non-essential parameters by decomposing weight matrices into products of two low-rank matrices. Yet, its application in LLMs has not been extensively studied. The key to low-rank compression lies in low-rank factorization and low-rank dimensions allocation. To address the challenges of low-rank compression in LLMs, we conduct empirical research on the low-rank characteristics of large models. We propose a low-rank compression method suitable for LLMs. This approach involves precise estimation of feature distributions through pooled covariance matrices and a Bayesian optimization strategy for allocating low-rank dimensions. Experiments on the LLaMA-2 models demonstrate that our method outperforms existing strong structured pruning and low-rank compression techniques in maintaining model performance at the same compression ratio.

Pengxiang Zhao, Ping Li, Yingjie Gu et al.

As deep learning models exponentially increase in size, optimizers such as Adam encounter significant memory consumption challenges due to the storage of first and second moment data. Current memory-efficient methods like Adafactor and CAME often compromise accuracy with their matrix factorization techniques. Addressing this, we introduce Adapprox, a novel approach that employs randomized low-rank matrix approximation for a more effective and accurate approximation of Adam's second moment. Adapprox features an adaptive rank selection mechanism, finely balancing accuracy and memory efficiency, and includes an optional cosine similarity guidance strategy to enhance stability and expedite convergence. In GPT-2 training and downstream tasks, Adapprox surpasses AdamW by achieving 34.5% to 49.9% and 33.8% to 49.9% memory savings for the 117M and 345M models, respectively, with the first moment enabled, and further increases these savings without the first moment. Besides, it enhances convergence speed and improves downstream task performance relative to its counterparts.

Ranran Zhen, Juntao Li, Yixin Ji et al.

Large Language Models (LLMs) for Generative AI have achieved remarkable progress, evolving into sophisticated and versatile tools widely adopted across various domains and applications. However, the substantial memory overhead caused by their vast number of parameters, combined with the high computational demands of the attention mechanism, poses significant challenges in achieving low latency and high throughput for LLM inference services. Recent advancements, driven by groundbreaking research, have significantly accelerated progress in this field. This paper provides a comprehensive survey of these methods, covering fundamental instance-level approaches, in-depth cluster-level strategies, emerging scenario directions, and other miscellaneous but important areas. At the instance level, we review model placement, request scheduling, decoding length prediction, storage management, and the disaggregation paradigm. At the cluster level, we explore GPU cluster deployment, multi-instance load balancing, and cloud service solutions. For emerging scenarios, we organize the discussion around specific tasks, modules, and auxiliary methods. To ensure a holistic overview, we also highlight several niche yet critical areas. Finally, we outline potential research directions to further advance the field of LLM inference serving.

Hanyu Hu, Pengxiang Zhao, Ping Li et al.

The rapid increase in the size of large language models (LLMs) has significantly escalated their computational and memory demands, posing challenges for efficient deployment, especially on resource-constrained devices. Structured pruning has emerged as an effective model compression method that can reduce these demands while preserving performance. In this paper, we introduce FASP (Fast and Accurate Structured Pruning), a novel structured pruning framework for LLMs that emphasizes both speed and accuracy. FASP employs a distinctive pruning structure that interlinks sequential layers, allowing for the removal of columns in one layer while simultaneously eliminating corresponding rows in the preceding layer without incurring additional performance loss. The pruning metric, inspired by Wanda, is computationally efficient and effectively selects components to prune. Additionally, we propose a restoration mechanism that enhances model fidelity by adjusting the remaining weights post-pruning. We evaluate FASP on the OPT and LLaMA model families, demonstrating superior performance in terms of perplexity and accuracy on downstream tasks compared to state-of-the-art methods. Our approach achieves significant speed-ups, pruning models such as OPT-125M in 17 seconds and LLaMA-30B in 15 minutes on a single NVIDIA RTX 4090 GPU, making it a highly practical solution for optimizing LLMs.

Feiyang Chen, Hongpeng Pan, Haonan Xu et al.

Diffusion distillation is central to accelerating image and video generation, yet existing methods are fundamentally limited by the denoising process, where step reduction has largely saturated. Partial timestep low-resolution generation can further accelerate inference, but it suffers noticeable quality degradation due to cross-resolution distribution gaps. We propose Cross-Resolution Distribution Matching Distillation (RMD), a novel distillation framework that bridges cross-resolution distribution gaps for high-fidelity, few-step multi-resolution cascaded inference. Specifically, RMD divides the timestep intervals for each resolution using logarithmic signal-to-noise ratio (logSNR) curves, and introduces logSNR-based mapping to compensate for resolution-induced shifts. Distribution matching is conducted along resolution trajectories to reduce the gap between low-resolution generator distributions and the teacher's high-resolution distribution. In addition, a predicted-noise re-injection mechanism is incorporated during upsampling to stabilize training and improve synthesis quality. Quantitative and qualitative results show that RMD preserves high-fidelity generation while accelerating inference across various backbones. Notably, RMD achieves up to 33.4X speedup on SDXL and 25.6X on Wan2.1-14B, while preserving high visual fidelity.

Jikai Wang, Zhenxu Tian, Juntao Li et al.

Recent works have revealed the great potential of speculative decoding in accelerating the autoregressive generation process of large language models. The success of these methods relies on the alignment between draft candidates and the sampled outputs of the target model. Existing methods mainly achieve draft-target alignment with training-based methods, e.g., EAGLE, Medusa, involving considerable training costs. In this paper, we present a training-free alignment-augmented speculative decoding algorithm. We propose alignment sampling, which leverages output distribution obtained in the prefilling phase to provide more aligned draft candidates. To further benefit from high-quality but non-aligned draft candidates, we also introduce a simple yet effective flexible verification strategy. Through an adaptive probability threshold, our approach can improve generation accuracy while further improving inference efficiency. Experiments on 8 datasets (including question answering, summarization and code completion tasks) show that our approach increases the average generation score by 3.3 points for the LLaMA3 model. Our method achieves a mean acceptance length up to 2.39 and speed up generation by 2.23.

Ping Chen, Zhuohong Deng, Ping Li et al.

Training large language models (LLMs) is often constrained by GPU memory limitations. To alleviate memory pressure, activation recomputation and data compression have been proposed as two major strategies. However, both approaches have limitations: recomputation introduces significant training overhead, while compression can lead to accuracy degradation and computational inefficiency when applied naively. In this paper, we propose Adacc, the first adaptive memory optimization framework that unifies activation recomputation and data compression to improve training efficiency for LLMs while preserving model accuracy. Unlike existing methods that apply static, rule-based strategies or rely solely on one technique, Adacc makes fine-grained, tensor-level decisions, dynamically selecting between recomputation, retention, and compression based on tensor characteristics and runtime hardware constraints. Adacc tackles three key challenges: (1) it introduces layer-specific compression algorithms that mitigate accuracy loss by accounting for outliers in LLM activations; (2) it employs a MILP-based scheduling policy to globally optimize memory strategies across layers; and (3) it integrates an adaptive policy evolution mechanism to update strategies during training in response to changing data distributions. Experimental results show that Adacc improves training throughput by 1.01x to 1.37x compared to state-of-the-art frameworks, while maintaining accuracy comparable to the baseline.

Yi Su, Quantong Qiu, Yuechi Zhou et al.

Large Language Models (LLMs) require substantial computational resources during generation. While the Key-Value (KV) cache significantly accelerates this process by storing attention intermediates, its memory footprint grows linearly with sequence length, batch size, and model size, creating a bottleneck in long-context scenarios. Various KV cache compression techniques, including token eviction, quantization, and low-rank projection, have been proposed to mitigate this bottleneck, often complementing each other. This paper focuses on enhancing token eviction strategies. Token eviction leverages the observation that the attention patterns are often sparse, allowing for the removal of less critical KV entries to save memory. However, this reduction usually comes at the cost of notable accuracy degradation, particularly under high compression ratios. To address this issue, we propose \textbf{CaliDrop}, a novel strategy that enhances token eviction through calibration. Our preliminary experiments show that queries at nearby positions exhibit high similarity. Building on this observation, CaliDrop performs speculative calibration on the discarded tokens to mitigate the accuracy loss caused by token eviction. Extensive experiments demonstrate that CaliDrop significantly improves the accuracy of existing token eviction methods.

Ping Chen, Wenjie Zhang, Shuibing He et al.

Large model training often uses recomputation to alleviate memory pressure and pipelines to exploit the parallelism of data, tensors, and devices. However, existing recomputation approaches may incur high overhead when training real-world models, as they are executed on demand in the critical training path. In this paper, we present Lynx, a new recomputation framework to reduce overhead by overlapping recomputation with communication in training pipelines. To reduce the large search space for recomputation strategies, we propose a heuristic-based recomputation scheduling algorithm, which is based on the observation that there are identical structures in large DNN models so that we can apply the same scheduling policy to all such structures. Additionally, we propose a recomputation-aware model partitioning method to balance each stage's execution time for improved training throughput. Our comprehensive evaluation using GPT models with 1.3B-23B parameters shows that Lynx outperforms existing recomputation approaches by up to 1.37x.

Shilin Zhou, Zhenghua Li, Yu Hong et al.

End-to-end automatic speech recognition (ASR) systems have made significant progress in general scenarios. However, it remains challenging to transcribe contextual named entities (NEs) in the contextual ASR scenario. Previous approaches have attempted to address this by utilizing the NE dictionary. These approaches treat entities as individual tokens and generate them token-by-token, which may result in incomplete transcriptions of entities. In this paper, we treat entities as indivisible wholes and introduce the idea of copying into ASR. We design a systematic mechanism called CopyNE, which can copy entities from the NE dictionary. By copying all tokens of an entity at once, we can reduce errors during entity transcription, ensuring the completeness of the entity. Experiments demonstrate that CopyNE consistently improves the accuracy of transcribing entities compared to previous approaches. Even when based on the strong Whisper, CopyNE still achieves notable improvements.

Abbas Ghaddar, Yimeng Wu, Ahmad Rashid et al.

Language-specific pre-trained models have proven to be more accurate than multilingual ones in a monolingual evaluation setting, Arabic is no exception. However, we found that previously released Arabic BERT models were significantly under-trained. In this technical report, we present JABER and SABER, Junior and Senior Arabic BERt respectively, our pre-trained language model prototypes dedicated for Arabic. We conduct an empirical study to systematically evaluate the performance of models across a diverse set of existing Arabic NLU tasks. Experimental results show that JABER and SABER achieve state-of-the-art performances on ALUE, a new benchmark for Arabic Language Understanding Evaluation, as well as on a well-established NER benchmark.

Rongjie Huang, Chenye Cui, Feiyang Chen et al.

Deep generative models have achieved significant progress in speech synthesis to date, while high-fidelity singing voice synthesis is still an open problem for its long continuous pronunciation, rich high-frequency parts, and strong expressiveness. Existing neural vocoders designed for text-to-speech cannot directly be applied to singing voice synthesis because they result in glitches and poor high-frequency reconstruction. In this work, we propose SingGAN, a generative adversarial network designed for high-fidelity singing voice synthesis. Specifically, 1) to alleviate the glitch problem in the generated samples, we propose source excitation with the adaptive feature learning filters to expand the receptive field patterns and stabilize long continuous signal generation; and 2) SingGAN introduces global and local discriminators at different scales to enrich low-frequency details and promote high-frequency reconstruction; and 3) To improve the training efficiency, SingGAN includes auxiliary spectrogram losses and sub-band feature matching penalty loss. To the best of our knowledge, SingGAN is the first work designed toward high-fidelity singing voice vocoding. Our evaluation of SingGAN demonstrates the state-of-the-art results with higher-quality (MOS 4.05) samples. Also, SingGAN enables a sample speed of 50x faster than real-time on a single NVIDIA 2080Ti GPU. We further show that SingGAN generalizes well to the mel-spectrogram inversion of unseen singers, and the end-to-end singing voice synthesis system SingGAN-SVS enjoys a two-stage pipeline to transform the music scores into expressive singing voices. Audio samples are available at \url{https://SingGAN.github.io/}

Chen Gong, Saihao Huang, Houquan Zhou et al.

Unlike English letters, Chinese characters have rich and specific meanings. Usually, the meaning of a word can be derived from its constituent characters in some way. Several previous works on syntactic parsing propose to annotate shallow word-internal structures for better utilizing character-level information. This work proposes to model the deep internal structures of Chinese words as dependency trees with 11 labels for distinguishing syntactic relationships. First, based on newly compiled annotation guidelines, we manually annotate a word-internal structure treebank (WIST) consisting of over 30K multi-char words from Chinese Penn Treebank. To guarantee quality, each word is independently annotated by two annotators and inconsistencies are handled by a third senior annotator. Second, we present detailed and interesting analysis on WIST to reveal insights on Chinese word formation. Third, we propose word-internal structure parsing as a new task, and conduct benchmark experiments using a competitive dependency parser. Finally, we present two simple ways to encode word-internal structures, leading to promising gains on the sentence-level syntactic parsing task.

Yingjie Gu, Xiaoye Qu, Zhefeng Wang et al.

Entity linking (EL) for the rapidly growing short text (e.g. search queries and news titles) is critical to industrial applications. Most existing approaches relying on adequate context for long text EL are not effective for the concise and sparse short text. In this paper, we propose a novel framework called Multi-turn Multiple-choice Machine reading comprehension (M3}) to solve the short text EL from a new perspective: a query is generated for each ambiguous mention exploiting its surrounding context, and an option selection module is employed to identify the golden entity from candidates using the query. In this way, M3 framework sufficiently interacts limited context with candidate entities during the encoding process, as well as implicitly considers the dissimilarities inside the candidate bunch in the selection stage. In addition, we design a two-stage verifier incorporated into M3 to address the commonly existed unlinkable problem in short text. To further consider the topical coherence and interdependence among referred entities, M3 leverages a multi-turn fashion to deal with mentions in a sequence manner by retrospecting historical cues. Evaluation shows that our M3 framework achieves the state-of-the-art performance on five Chinese and English datasets for the real-world short text EL.

Yang Li, Jianke Zhu, Steven C. H. Hoi et al.

Most of existing correlation filter-based tracking approaches only estimate simple axis-aligned bounding boxes, and very few of them is capable of recovering the underlying similarity transformation. To tackle this challenging problem, in this paper, we propose a new correlation filter-based tracker with a novel robust estimation of similarity transformation on the large displacements. In order to efficiently search in such a large 4-DoF space in real-time, we formulate the problem into two 2-DoF sub-problems and apply an efficient Block Coordinates Descent solver to optimize the estimation result. Specifically, we employ an efficient phase correlation scheme to deal with both scale and rotation changes simultaneously in log-polar coordinates. Moreover, a variant of correlation filter is used to predict the translational motion individually. Our experimental results demonstrate that the proposed tracker achieves very promising prediction performance compared with the state-of-the-art visual object tracking methods while still retaining the advantages of high efficiency and simplicity in conventional correlation filter-based tracking methods.

Lingyang Chu, Zhefeng Wang, Jian Pei et al.

Given a database network where each vertex is associated with a transaction database, we are interested in finding theme communities. Here, a theme community is a cohesive subgraph such that a common pattern is frequent in all transaction databases associated with the vertices in the subgraph. Finding all theme communities from a database network enjoys many novel applications. However, it is challenging since even counting the number of all theme communities in a database network is #P-hard. Inspired by the observation that a theme community shrinks when the length of the pattern increases, we investigate several properties of theme communities and develop TCFI, a scalable algorithm that uses these properties to effectively prune the patterns that cannot form any theme community. We also design TC-Tree, a scalable algorithm that decomposes and indexes theme communities efficiently. Retrieving a ranked list of theme communities from a TC-Tree of hundreds of millions of theme communities takes less than 1 second. Extensive experiments and a case study demonstrate the effectiveness and scalability of TCFI and TC-Tree in discovering and querying meaningful theme communities from large database networks.