Manyi Zhang, Ji-Fu Li, Zhongao Sun, Haoli Bai, Hui-Ling Zhen, Zhenhua Dong, Xianzhi Yu. Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2026.

GPT-4o, an omni-modal model that enables vocal conversations with diverse emotions and tones, marks a milestone for omni-modal foundation models. However, empowering Large Language Models to perceive and generate images, texts, and speeches end-to-end with publicly available data remains challenging for the open-source community. Existing vision-language models rely on external tools for speech processing, while speech-language models still suffer from limited or totally without vision-understanding capabilities. To address this gap, we propose the EMOVA (EMotionally Omni-present Voice Assistant), to enable Large Language Models with end-to-end speech abilities while maintaining the leading vision-language performance. With a semantic-acoustic disentangled speech tokenizer, we surprisingly notice that omni-modal alignment can further enhance vision-language and speech abilities compared with the bi-modal aligned counterparts. Moreover, a lightweight style module is introduced for the flexible speech style controls including emotions and pitches. For the first time, EMOVA achieves state-of-the-art performance on both the vision-language and speech benchmarks, and meanwhile, supporting omni-modal spoken dialogue with vivid emotions.

Large Language Models (LLMs) use key-value (KV) cache to reduce redundant computation in autoregressive generation. However, the KV cache size increases linearly during generation, leading to excessive memory usage, especially for long texts. Most KV cache compression methods evict the unimportant KV pairs to maintain a fixed cache size, which leads to the permanent loss of tokens during generation. However, singular value decomposition shows that values do not exhibit a strong low-rank property as keys do, suggesting that information is distributed more evenly across values, in contrast to its more redundant distribution within keys. Therefore, methods that evict both keys and values risk losing crucial information and compromise context integrity, ultimately degrading the output quality. To address this problem, we propose WeightedKV, a novel, training-free approach that discards the keys of less important tokens, while merging their values into neighboring tokens via a convex combination weighted by their average attention scores. In this way, the retained keys serve as anchors that guide the generation process, while the merged values provide a rich contextual backdrop. We assess our method on four widely used language modeling datasets, demonstrating superior performance compared to all baseline methods, particularly with a lower budget ratio.

Recently, quantization has been widely used for the compression and acceleration of large language models (LLMs). Due to the outliers in LLMs, it is crucial to flatten weights and activations to minimize quantization error with equally spaced quantization points. Prior research explores various pre-quantization transformations to suppress outliers, such as per-channel scaling and Hadamard transformation. However, we observe that these transformed weights and activations can still exhibit steep and dispersed distributions. In this paper, we propose FlatQuant (Fast and Learnable Affine Transformation), a new post-training quantization approach that enhances the flatness of weights and activations. Our approach identifies optimal affine transformations for each linear layer, calibrated in hours via a lightweight objective. To reduce runtime overhead of affine transformation, we apply Kronecker product with two lightweight matrices, and fuse all operations in FlatQuant into a single kernel. Extensive experiments demonstrate that FlatQuant establishes a new state-of-the-art benchmark for quantization. For example, it achieves less than 1\% accuracy drop for W4A4 quantization on the LLaMA-3-70B model, surpassing SpinQuant by 7.5\%. Additionally, it provides up to 2.3x prefill speedup and 1.7x decoding speedup compared to the FP16 model. Code is available at: https://github.com/ruikangliu/FlatQuant.

Efficient key-value (KV) cache compression is critical for scaling transformer-based Large Language Models (LLMs) in long sequences and resource-limited settings. Existing methods evict tokens based on their positions or importance, but position-based strategies can miss crucial information outside predefined regions, while those relying on global importance scores resulting in strong regional biases, limiting the KV cache's overall context retention and potentially impairing the performance of LLMs on complex tasks. Our wavelet analysis reveals that as tokens approach the end of sequence, their contributions to generation gradually increase and tends to diverge more from neighboring tokens, indicating a smooth transition with increasing complexity and variability from distant to nearby context. Motivated by this observation, we propose TreeKV, an intuitive, training-free method that employs a tree structure for smooth cache compression. TreeKV maintains a fixed cache size, allowing LLMs to deliver high-quality output in long text scenarios and is applicable during both the generation and prefilling stages. TreeKV consistently surpasses all baseline models in language modeling tasks on PG19 and OpenWebText2, allowing LLMs trained with short context window to generalize to longer window with a 16x cache reduction. On the Longbench benchmark, TreeKV achieves the best performance with only 6% of the budget at optimal efficiency.

Layer pruning has emerged as a widely used technique for compressing large language models (LLMs). However, existing layer pruning approaches often incur substantial performance degradation. We identify the majority of this degradation to a single yet previously overlooked issue: \textit{the mismatch of activation magnitudes at the pruning interface}. The pre-interface activations exhibit significantly different scales from the post-interface ones, causing the distributional shift as it propagates through the remaining layers. To address this issue, we introduce \textsc{LinearPatch}, a lightweight and plug-and-play technique that fuses two operations into one matrix multiply at the pruning interface: (i) a Hadamard transformation that suppresses massive outliers at particular tokens and (ii) a channel-wise scaling that aligns activation statistics. On LLaMA-3-8B, \textsc{LinearPatch} preserves up to \textbf{94.15\%} of the original model's performance when pruning 5 out of 32 layers, outperforming the previous state of the art by \textbf{4\%}. The patch can be further refined with 5K unlabeled samples via memory-efficient offline distillation, pushing the retention to 95.16\% within only 30 minutes on a single GPU. Code is available at \url{https://github.com/chenxinrui-tsinghua/LinearPatch}.

Vision-language pretrained models have achieved impressive performance on various downstream tasks. However, their large model sizes hinder their utilization on platforms with limited computational resources. We find that directly using smaller pretrained models and applying magnitude-based pruning on CLIP models leads to in-flexibility and inferior performance. Recent efforts for VLP compression either adopt uni-modal compression metrics resulting in limited performance or involve costly mask-search processes with learnable masks. In this paper, we first propose the Module-wise Pruning Error (MoPE) met-ric, accurately assessing CLIP module importance by performance decline on cross-modal tasks. Using the MoPE metric, we introduce a unified pruning framework applica-ble to both pretraining and task-specific fine-tuning compression stages. For pretraining, MoPE-CLIP effectively leverages knowledge from the teacher model, significantly reducing pretraining costs while maintaining strong zero-shot capabilities. For fine-tuning, consecutive pruning from width to depth yields highly competitive task-specific models. Extensive experiments in two stages demonstrate the effectiveness of the MoPE metric, and MoPE-CLIP outperforms previous state-of-the-art VLP compression methods.

With the rapid growth of large language models (LLMs), there is increasing demand for memory and computation in LLMs. Recent efforts on post-training pruning of LLMs aim to reduce the model size and computation requirements, yet the performance is still sub-optimal. In this paper, we present a plug-and-play solution for post-training pruning of LLMs. The proposed solution has two innovative components: 1) **Relative Importance and Activations (RIA)**, a new pruning metric that jointly considers the weight and activations efficiently on LLMs, and 2) **Channel Permutation**, a new approach to maximally preserves important weights under N:M sparsity. The two proposed components can be readily combined to further enhance the N:M semi-structured pruning of LLMs. Our empirical experiments show that RIA alone can already surpass all existing post-training pruning methods on prevalent LLMs, e.g., LLaMA ranging from 7B to 65B. Furthermore, N:M semi-structured pruning with channel permutation can even outperform the original LLaMA2-70B on zero-shot tasks, together with practical speed-up on specific hardware. Our code is available at: https://github.com/biomedical-cybernetics/Relative-importance-and-activation-pruning

Zhiming Mao, Haoli Bai, Lu Hou, Lifeng Shang, Xin Jiang, Qun Liu, Kam-Fai Wong. Proceedings of the 2024 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers). 2024.

Haoli Bai, Zhiguang Liu, Xiaojun Meng, Li Wentao, Shuang Liu, Yifeng Luo, Nian Xie, Rongfu Zheng, Liangwei Wang, Lu Hou, Jiansheng Wei, Xin Jiang, Qun Liu. Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2023.

As one of the successful Transformer-based models in computer vision tasks, SegFormer demonstrates superior performance in semantic segmentation. Nevertheless, the high computational cost greatly challenges the deployment of SegFormer on edge devices. In this paper, we seek to design a lightweight SegFormer for efficient semantic segmentation. Based on the observation that neurons in SegFormer layers exhibit large variances across different images, we propose a dynamic gated linear layer, which prunes the most uninformative set of neurons based on the input instance. To improve the dynamically pruned SegFormer, we also introduce two-stage knowledge distillation to transfer the knowledge within the original teacher to the pruned student network. Experimental results show that our method can significantly reduce the computation overhead of SegFormer without an apparent performance drop. For instance, we can achieve 36.9% mIoU with only 3.3G FLOPs on ADE20K, saving more than 60% computation with the drop of only 0.5% in mIoU.

Network quantization has gained increasing attention with the rapid growth of large pre-trained language models~(PLMs). However, most existing quantization methods for PLMs follow quantization-aware training~(QAT) that requires end-to-end training with full access to the entire dataset. Therefore, they suffer from slow training, large memory overhead, and data accessibility issues. In this paper, we study post-training quantization~(PTQ) of PLMs, and propose module-wise quantization error minimization~(MREM), an efficient solution to mitigate these issues. By partitioning the PLM into multiple modules, we minimize the reconstruction error incurred by quantization for each module. In addition, we design a new model parallel training strategy such that each module can be trained locally on separate computing devices without waiting for preceding modules, which brings nearly the theoretical training speed-up (e.g., $4\times$ on $4$ GPUs). Experiments on GLUE and SQuAD benchmarks show that our proposed PTQ solution not only performs close to QAT, but also enjoys significant reductions in training time, memory overhead, and data consumption.

Haoli Bai, Wei Zhang, Lu Hou, Lifeng Shang, Jin Jin, Xin Jiang, Qun Liu, Michael Lyu, Irwin King. Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers). 2021.

Variational autoencoders (VAEs) have been widely applied for text modeling. In practice, however, they are troubled by two challenges: information underrepresentation and posterior collapse. The former arises as only the last hidden state of LSTM encoder is transformed into the latent space, which is generally insufficient to summarize the data. The latter is a longstanding problem during the training of VAEs as the optimization is trapped to a disastrous local optimum. In this paper, we propose Discrete Auto-regressive Variational Attention Model (DAVAM) to address the challenges. Specifically, we introduce an auto-regressive variational attention approach to enrich the latent space by effectively capturing the semantic dependency from the input. We further design discrete latent space for the variational attention and mathematically show that our model is free from posterior collapse. Extensive experiments on language modeling tasks demonstrate the superiority of DAVAM against several VAE counterparts. Code will be released.

Model compression has been widely adopted to obtain light-weighted deep neural networks. Most prevalent methods, however, require fine-tuning with sufficient training data to ensure accuracy, which could be challenged by privacy and security issues. As a compromise between privacy and performance, in this paper we investigate few shot network compression: given few samples per class, how can we effectively compress the network with negligible performance drop? The core challenge of few shot network compression lies in high estimation errors from the original network during inference, since the compressed network can easily over-fits on the few training instances. The estimation errors could propagate and accumulate layer-wisely and finally deteriorate the network output. To address the problem, we propose cross distillation, a novel layer-wise knowledge distillation approach. By interweaving hidden layers of teacher and student network, layer-wisely accumulated estimation errors can be effectively reduced. The proposed method offers a general framework compatible with prevalent network compression techniques such as pruning. Extensive experiments n benchmark datasets demonstrate that cross distillation can significantly improve the student network's accuracy when only a few training instances are available.

To deploy deep neural networks on resource-limited devices, quantization has been widely explored. In this work, we study the extremely low-bit networks which have tremendous speed-up, memory saving with quantized activation and weights. We first bring up three omitted issues in extremely low-bit networks: the squashing range of quantized values; the gradient vanishing during backpropagation and the unexploited hardware acceleration of ternary networks. By reparameterizing quantized activation and weights vector with full precision scale and offset for fixed ternary vector, we decouple the range and magnitude from direction to extenuate above problems. Learnable scale and offset can automatically adjust the range of quantized values and sparsity without gradient vanishing. A novel encoding and computation pattern are designed to support efficient computing for our reparameterized ternary network (RTN). Experiments on ResNet-18 for ImageNet demonstrate that the proposed RTN finds a much better efficiency between bitwidth and accuracy and achieves up to 26.76% relative accuracy improvement compared with state-of-the-art methods. Moreover, we validate the proposed computation pattern on Field Programmable Gate Arrays (FPGA), and it brings 46.46 × and 89.17 × savings on power and area compared with the full precision convolution.

Neural Architecture Search (NAS) has recently outperformed hand-crafted networks in various areas. However, most prevalent NAS methods only focus on a pre-defined task. For a previously unseen task, the architecture is either searched from scratch, which is inefficient, or transferred from the one obtained on some other task, which might be sub-optimal. In this paper, we investigate a previously unexplored problem: whether a universal NAS method exists, such that task-aware architectures can be effectively generated? Towards this problem, we propose Meta Neural Architecture Search (M-NAS). To obtain task-specific architectures, M-NAS adopts a task-aware architecture controller for child model generation. Since optimal weights for different tasks and architectures span diversely, we resort to meta-learning, and learn meta-weights that efficiently adapt to a new task on the corresponding architecture with only several gradient descent steps. Experimental results demonstrate the superiority of M-NAS against a number of competitive baselines on both toy regression and few shot classification problems.

In transfer learning, what and where to transfer has been widely studied. Nevertheless, the learned transfer strategies are at high risk of over-fitting, especially when only a few annotated instances are available in the target domain. In this paper, we introduce the concept of transfer ensemble learning, a new direction to tackle the over-fitting of transfer strategies. Intuitively, models with different transfer strategies offer various perspectives on what and where to transfer. Therefore a core problem is to search these diversely transferred models for ensemble so as to achieve better generalization. Towards this end, we propose the Transferability Slider (TranSlider) for transfer ensemble learning. By decreasing the transferability, we obtain a spectrum of base models ranging from pure exploitation of the source model to unconstrained exploration for the target domain. Furthermore, the manner of decreasing transferability with parameter sharing guarantees fast optimization at no additional training cost. Finally, we conduct extensive experiments with various analyses, which demonstrate that TranSlider achieves the state-of-the-art on comprehensive benchmark datasets.

Recent advances in neural architecture search inspire many channel number search algorithms~(CNS) for convolutional neural networks. To improve searching efficiency, parameter sharing is widely applied, which reuses parameters among different channel configurations. Nevertheless, it is unclear how parameter sharing affects the searching process. In this paper, we aim at providing a better understanding and exploitation of parameter sharing for CNS. Specifically, we propose affine parameter sharing~(APS) as a general formulation to unify and quantitatively analyze existing channel search algorithms. It is found that with parameter sharing, weight updates of one architecture can simultaneously benefit other candidates. However, it also results in less confidence in choosing good architectures. We thus propose a new strategy of parameter sharing towards a better balance between training efficiency and architecture discrimination. Extensive analysis and experiments demonstrate the superiority of the proposed strategy in channel configuration against many state-of-the-art counterparts on benchmark datasets.

Link prediction is a fundamental problem in network modeling. A family of link prediction approaches is to treat network data as an exchangeable array whose entries can be explained by random functions (e.g., block models and Gaussian processes) over latent node factors. Despite their powerful ability in modeling missing links, these models tend to have a large computational complexity and thus are hard to deal with large networks. To address this problem, we develop a novel variational random function model by defining latent Gaussian processes on exchangeable arrays. This model not only inherits the ability of Gaussian process to describe the nonlinear interactions between nodes, but also enjoys significant reduction on computational complexity. To further make the model scalable to large network data, we develop an efficient key-value-free strategy under the map-reduce framework to tremendously reduce the inference time. Experimental results on large network data have demonstrated both the efficacy and efficiency of the proposed method over state-of-the-arts methods in network modeling.

Topic modelling and citation recommendation of scientific articles are important yet challenging research problems in scientific article analysis. In particular, the inference on coherent topics can be easily affected by irrelevant contents in articles. Meanwhile, the extreme sparsity of citation networks brings difficulty to a valid citation recommendation. Intuitively, articles with similar topics are more likely to cite each other, and cited articles tend to share similar themes. Motivated from this intuition, we aim to boost the performance of both topic modelling and citation recommendation by effectively leverage this underlying correlation between latent topics and citation networks. To this end, we propose a novel Bayesian deep generative model termed as Neural Relational Topic Model (NRTM), which is composed with a Stacked Variational Auto-Encoder (SVAE) and a multilayer perception (MLP). Specifically, the SVAE utilizes an inference network to learn more representative topics of document contents, which can help to enrich the latent factors in collaborative filtering of citations. Furthermore, the MLP network conducts nonlinear collaborative filtering of citations, which can further benefit the inference of topics by leveraging the knowledge of citation networks. Extensive experiments on two real-world datasets demonstrate that our model can effectively take advantages of the coherence between topic learning and citation recommendation, and significantly outperform the state-of-the-art methods on both tasks.

Segmentation and labeling for high dimensional time series is an important yet challenging task in a number of applications, such as behavior understanding and medical diagnosis. Recent advances to model the nonlinear dynamics in such time series data, has suggested to involve recurrent neural networks into Hidden Markov Models. However, this involvement has caused the inference procedure much more complicated, often leading to intractable inference, especially for the discrete variables of segmentation and labeling. To achieve both flexibility and tractability in modeling nonlinear dynamics of discrete variables, we present a structured and stochastic sequential neural network (SSNN), which composes with a generative network and an inference network. In detail, the generative network aims to not only capture the long-term dependencies but also model the uncertainty of the segmentation labels via semi-Markov models. More importantly, for efficient and accurate inference, the proposed bi-directional inference network reparameterizes the categorical segmentation with the Gumbel-Softmax approximation and resorts to the Stochastic Gradient Variational Bayes. We evaluate the proposed model in a number of tasks, including speech modeling, automatic segmentation and labeling in behavior understanding, and sequential multi-objects recognition. Experimental results have demonstrated that our proposed model can achieve significant improvement over the state-of-the-art methods.

We consider a different setting from regular multitask learning, where data from different tasks share no common instances and no common feature dictionary, while the features can be semantically correlated and all tasks share the same class space. For example, in the two tasks of identifying terrorism information from English news and Arabic news respectively, one associated dataset could be news from Cable News Network (CNN), and the other be news crawled from websites of Arabic countries. Intuitively, these two tasks could help each other, although they share no common feature space. This new setting has brought obstacles to traditional multi-task learning algorithms and multi-view learning algorithms. We argue that these different data sources can be co-trained together by exploring the latent semantics among them. To this end, we propose a new graphical model based on sparse Gaussian Conditional Random Fields (GCRF) and Hilbert-Schmidt Independence Criterion (HSIC). In additional to output the prediction accuracy for each single task, it can also model (1) the dependency between the latent feature spaces of different tasks, (2) the dependency of the category spaces, and (3) the dependency between the latent feature space and the category space in each task. To make the model inference effective, we have provided an efficient variational EM algorithm. Experiments on both synthetic data sets and real-world data sets have indicated the feasibility and effectiveness of the proposed framework.

Link prediction in multi-relational social networks has attracted much attention. For instance, we may care the chance of two users being friends based on their contacts of other patterns, e.g., SMS and phone calls. In previous work, matrix factorization models are typically applied in single-relational networks; however, two challenges arise to extend it into multi-relational networks. First, the interaction of different relation types is hard to be captured. The second is the cold start problem, as the prediction of new entities in multi-relational networks becomes even more challenging. In this article we propose a novel method called Hierarchical Probabilistic Matrix Factorization with Network Topology (HPMFNT). Our model exploits the network topology by extending the Katz index into multi-relational settings, which could efficiently model the multidimensional interplay via the auxiliary information from other relationships. We also utilize the extended Katz index along with entitiy attributes to solve the cold-start problem. Experiments on two real world datasets have shown that our model outperforms the state-of-the-art with a significant margin.