Hengyuan Zhang, Shiping Yang, Xiao Liang, Chenming Shang, Yuxuan Jiang, Chaofan Tao, Jing Xiong, Hayden Kwok-Hay So, Ruobing Xie, Angel X. Chang, Ngai Wong. Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2026.

Kullback-Leiber divergence has been widely used in Knowledge Distillation (KD) to compress Large Language Models (LLMs). Contrary to prior assertions that reverse Kullback-Leibler (RKL) divergence is mode-seeking and thus preferable over the mean-seeking forward Kullback-Leibler (FKL) divergence, this study empirically and theoretically demonstrates that neither mode-seeking nor mean-seeking properties manifest in KD for LLMs. Instead, RKL and FKL are found to share the same optimization objective and both converge after a sufficient number of epochs. However, due to practical constraints, LLMs are seldom trained for such an extensive number of epochs. Meanwhile, we further find that RKL focuses on the tail part of the distributions, while FKL focuses on the head part at the beginning epochs. Consequently, we propose a simple yet effective Adaptive Kullback-Leiber (AKL) divergence method, which adaptively allocates weights to combine FKL and RKL. Metric-based and GPT-4-based evaluations demonstrate that the proposed AKL outperforms the baselines across various tasks and improves the diversity and quality of generated responses.

Jing Xiong, Jianghan Shen, Fanghua Ye, Chaofan Tao, Zhongwei Wan, Jianqiao Lu, Xun Wu, Chuanyang Zheng, Zhijiang Guo, Min Yang, Lingpeng Kong, Ngai Wong. Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing. 2025.

Multimodal large language models (MLLMs) have shown promising capabilities in reasoning tasks, yet still struggle significantly with complex problems requiring explicit self-reflection and self-correction, especially compared to their unimodal text-based counterparts. Existing reflection methods are simplistic and struggle to generate meaningful, instructive feedback, as the reasoning ability and knowledge limits of pre-trained models are largely fixed during initial training. To overcome these challenges, we propose \textit{multimodal \textbf{S}elf-\textbf{R}eflection enhanced reasoning with Group Relative \textbf{P}olicy \textbf{O}ptimization} \textbf{SRPO}, a two-stage reflection-aware reinforcement learning (RL) framework explicitly designed to enhance multimodal LLM reasoning. In the first stage, we construct a high-quality, reflection-focused dataset under the guidance of an advanced MLLM, which generates reflections based on initial responses to help the policy model to learn both reasoning and self-reflection. In the second stage, we introduce a novel reward mechanism within the GRPO framework that encourages concise and cognitively meaningful reflection while avoiding redundancy. Extensive experiments across multiple multimodal reasoning benchmarks—including MathVista, MathVision, Mathverse, and MMMU-Pro—using Qwen-2.5-VL-7B and Qwen-2.5-VL-32B demonstrate that SRPO significantly outperforms state-of-the-art models, achieving notable improvements in both reasoning accuracy and reflection quality.

Robotic behavior synthesis, the problem of understanding multimodal inputs and generating precise physical control for robots, is an important part of Embodied AI. Despite successes in applying multimodal large language models for high-level understanding, it remains challenging to translate these conceptual understandings into detailed robotic actions while achieving generalization across various scenarios. In this paper, we propose a tree-structured multimodal code generation framework for generalized robotic behavior synthesis, termed RoboCodeX. RoboCodeX decomposes high-level human instructions into multiple object-centric manipulation units consisting of physical preferences such as affordance and safety constraints, and applies code generation to introduce generalization ability across various robotics platforms. To further enhance the capability to map conceptual and perceptual understanding into control commands, a specialized multimodal reasoning dataset is collected for pre-training and an iterative self-updating methodology is introduced for supervised fine-tuning. Extensive experiments demonstrate that RoboCodeX achieves state-of-the-art performance in both simulators and real robots on four different kinds of manipulation tasks and one embodied navigation task.

Recent vision-language models have achieved tremendous advances. However, their computational costs are also escalating dramatically, making model acceleration exceedingly critical. To pursue more efficient vision-language Transformers, this paper introduces Cross-Guided Ensemble of Tokens (CrossGET), a general acceleration framework for vision-language Transformers. This framework adaptively combines tokens in real-time during inference, significantly reducing computational costs while maintaining high performance. CrossGET features two primary innovations: 1) Cross-Guided Matching and Ensemble. CrossGET leverages cross-modal guided token matching and ensemble to effectively utilize cross-modal information, achieving wider applicability across both modality-independent models, e.g., CLIP, and modality-dependent ones, e.g., BLIP2. 2) Complete-Graph Soft Matching. CrossGET introduces an algorithm for the token-matching mechanism, ensuring reliable matching results while facilitating parallelizability and high efficiency. Extensive experiments have been conducted on various vision-language tasks, such as image-text retrieval, visual reasoning, image captioning, and visual question answering. The performance on both classic multimodal architectures and emerging multimodal LLMs demonstrates the framework’s effectiveness and versatility. The code is available at https://github.com/sdc17/CrossGET.

Research on scaling large language models (LLMs) has primarily focused on model parameters and training data size, overlooking the role of vocabulary size. We investigate how vocabulary size impacts LLM scaling laws by training models ranging from 33M to 3B parameters on up to 500B characters with various vocabulary configurations. We propose three complementary approaches for predicting the compute-optimal vocabulary size: IsoFLOPs analysis, derivative estimation, and parametric fit of the loss function. Our approaches converge on the conclusion that the optimal vocabulary size depends on the compute budget, with larger models requiring larger vocabularies. Most LLMs, however, use insufficient vocabulary sizes. For example, we predict that the optimal vocabulary size of Llama2-70B should have been at least 216K, 7 times larger than its vocabulary of 32K. We validate our predictions empirically by training models with 3B parameters across different FLOPs budgets. Adopting our predicted optimal vocabulary size consistently improves downstream performance over commonly used vocabulary sizes. By increasing the vocabulary size from the conventional 32K to 43K, we improve performance on ARC-Challenge from 29.1 to 32.0 with the same 2.3e21 FLOPs. Our work highlights the importance of jointly considering tokenization and model scaling for efficient pre-training. The code and demo are available at https://github.com/sail-sg/scaling-with-vocab and https://hf.co/spaces/sail/scaling-with-vocab-demo.

To accelerate the inference of deep neural networks (DNNs), quantization with low-bitwidth numbers is actively researched. A prominent challenge is to quantize the DNN models into low-bitwidth numbers without significant accuracy degradation, especially at very low bitwidths (< 8 bits). This work targets an adaptive data representation with variablelength encoding called DyBit. DyBit can dynamically adjust the precision and range of separate bit-fields to be adapted to the DNN weights/activations distribution. We also propose a hardware-aware quantization framework with a mixed-precision accelerator to trade-off the inference accuracy and speedup. Experimental results demonstrate that the ImageNet inference accuracy via DyBit is 1.97% higher than the state-of-the-art at 4-bit quantization, and the proposed framework can achieve up to 8.1× speedup compared with the original ResNet-50 model.

Learning low-bitwidth convolutional neural networks (CNNs) is challenging because performance may drop significantly after quantization. Prior arts often quantize the network weights by carefully tuning hyperparameters such as nonuniform stepsize and layerwise bitwidths, which are complicated since the full- and low-precision representations have large discrepancies. This work presents a novel quantization pipeline, named frequency-aware transformation (FAT), that features important benefits: 1) instead of designing complicated quantizers, FAT learns to transform network weights in the frequency domain to remove redundant information before quantization, making them amenable to training in low bitwidth with simple quantizers; 2) FAT readily embeds CNNs in low bitwidths using standard quantizers without tedious hyperparameter tuning and theoretical analyses show that FAT minimizes the quantization errors in both uniform and nonuniform quantizations; and 3) FAT can be easily plugged into various CNN architectures. Using FAT with a simple uniform/logarithmic quantizer can achieve the state-of-the-art performance in different bitwidths on various model architectures. Consequently, FAT serves to provide a novel frequency-based perspective for model quantization.

Real-world data contains a vast amount of multimodal information, among which vision and language are the two most representative modalities. Moreover, increasingly heavier models, e.g., Transformers, have attracted the attention of researchers to model compression. However, how to compress multimodal models, especially vison-language Transformers, is still under-explored. This paper proposes the Unified and Progressive Pruning (UPop) as a universal vison-language Transformer compression framework, which incorporates 1) unifiedly searching multimodal subnets in a continuous optimization space from the original model, which enables automatic assignment of pruning ratios among compressible modalities and structures; 2) progressively searching and retraining the subnet, which maintains convergence between the search and retrain to attain higher compression ratios. Experiments on various tasks, datasets, and model architectures demonstrate the effectiveness and versatility of the proposed UPop framework. The code is available at https://github.com/sdc17/UPop.

Chaofan Tao, Lu Hou, Wei Zhang, Lifeng Shang, Xin Jiang, Qun Liu, Ping Luo, Ngai Wong. Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2022.

Recent large-scale video-language pre-trained models have shown appealing performance on various downstream tasks.However, the pretraining process is computationally expensive due to the requirement of millions of videotext pairs and the redundant data structure of each video.To mitigate these problems, we propose LiteVL, which adapts a pre-trained image-language model BLIP into a video-text model directly on downstream tasks, without heavy pre-training.To enhance the temporal modeling lacking in the image-language model, we propose to add temporal attention modules in the image encoder of BLIP with dynamic temporal scaling.Besides the model-wise adaptation, we also propose a non-parametric pooling mechanism to adaptively reweight the fine-grained video embedding conditioned on the text.Experimental results on text-video retrieval and video question answering show that the proposed LiteVL even outperforms previous video-language pre-trained models by a clear margin, though without any videolanguage pre-training.

Quantizing neural networks to low-bitwidth is important for model deployment on resource-limited edge hardware. Although a quantized network has a smaller model size and memory footprint, it is fragile to adversarial attacks. However, few methods study the robustness and training efficiency of quantized networks. To this end, we propose a new method by recasting robust quantization as an online domain generalization problem, termed ODG-Q, which generates diverse adversarial data at a low cost during training. ODG-Q consistently outperforms existing works against various adversarial attacks. For example, on CIFAR-10 dataset, ODG-Q achieves 49.2% average improvements under five common white-box attacks and 21.7% average improvements under five common black-box attacks, with a training cost similar to that of natural training (viz. without adversaries). To our best knowledge, this work is the first work that trains both quantized and binary neural networks on ImageNet that consistently improves robustness under different attacks. We also provide a theoretical insight of ODG-Q that accounts for the bound of model risk on attacked data.

Transformers have shown great potential for modeling long-term dependencies for natural language processing and computer vision. However, little study has applied transformers to graphs, which is challenging due to the poor scalability of the attention mechanism and the under-exploration of graph inductive bias. To bridge this gap, we propose a Lite Graph Transformer (LiteGT) that learns on arbitrary graphs efficiently. First, a node sampling strategy is proposed to sparsify the considered nodes in self-attention with only O (Nlog N) time. Second, we devise two kernelization approaches to form two-branch attention blocks, which not only leverage graph-specific topology information, but also reduce computation further to O (1 over 2 Nlog N). Third, the nodes are updated with different attention schemes during training, thus largely mitigating over-smoothing problems when the model layers deepen. Extensive experiments demonstrate that LiteGT achieves competitive performance on both node classification and link prediction on datasets with millions of nodes. Specifically, Jaccard + Sampling + Dim. reducing setting reduces more than 100x computation and halves the model size without performance degradation.

Inferring the interactions between objects, a.k.a visual relationship detection, is a crucial point for vision understanding, which captures more definite concepts than object detection. Most previous work that treats the interaction between a pair of objects as a one way fail to exploit the mutual relation between objects, which is essential to modern visual application. In this work, we propose a mutual relation net, dubbed MR-Net, to explore the mutual relation between paired objects for visual relationship detection. Specifically, we construct a mutual relation space to model the mutual interaction of paired objects, and employ linear constraint to optimize the mutual interaction, which is called mutual relation learning. Our mutual relation learning does not introduce any parameters, and can adapt to improve the performance of other methods. In addition, we devise a semantic ranking loss to discriminatively penalize predicates with semantic similarity, which is ignored by traditional loss function (e.g., cross entropy with softmax). Then, our MR-Net optimizes the mutual relation learning together with semantic ranking loss with a siamese network. The experimental results on two commonly used datasets (VG and VRD) demonstrate the superior performance of the proposed approach.