Liang Song

Aiyuan Yang, Bin Xiao, Bingning Wang et al. · pku

Large language models (LLMs) have demonstrated remarkable performance on a variety of natural language tasks based on just a few examples of natural language instructions, reducing the need for extensive feature engineering. However, most powerful LLMs are closed-source or limited in their capability for languages other than English. In this technical report, we present Baichuan 2, a series of large-scale multilingual language models containing 7 billion and 13 billion parameters, trained from scratch, on 2.6 trillion tokens. Baichuan 2 matches or outperforms other open-source models of similar size on public benchmarks like MMLU, CMMLU, GSM8K, and HumanEval. Furthermore, Baichuan 2 excels in vertical domains such as medicine and law. We will release all pre-training model checkpoints to benefit the research community in better understanding the training dynamics of Baichuan 2.

Yang Liu, Dingkang Yang, Yan Wang et al.

Video Anomaly Detection (VAD) serves as a pivotal technology in the intelligent surveillance systems, enabling the temporal or spatial identification of anomalous events within videos. While existing reviews predominantly concentrate on conventional unsupervised methods, they often overlook the emergence of weakly-supervised and fully-unsupervised approaches. To address this gap, this survey extends the conventional scope of VAD beyond unsupervised methods, encompassing a broader spectrum termed Generalized Video Anomaly Event Detection (GVAED). By skillfully incorporating recent advancements rooted in diverse assumptions and learning frameworks, this survey introduces an intuitive taxonomy that seamlessly navigates through unsupervised, weakly-supervised, supervised and fully-unsupervised VAD methodologies, elucidating the distinctions and interconnections within these research trajectories. In addition, this survey facilitates prospective researchers by assembling a compilation of research resources, including public datasets, available codebases, programming tools, and pertinent literature. Furthermore, this survey quantitatively assesses model performance, delves into research challenges and directions, and outlines potential avenues for future exploration.

Kun Yang, Dingkang Yang, Jingyu Zhang et al.

Multi-agent collaborative perception as a potential application for vehicle-to-everything communication could significantly improve the perception performance of autonomous vehicles over single-agent perception. However, several challenges remain in achieving pragmatic information sharing in this emerging research. In this paper, we propose SCOPE, a novel collaborative perception framework that aggregates the spatio-temporal awareness characteristics across on-road agents in an end-to-end manner. Specifically, SCOPE has three distinct strengths: i) it considers effective semantic cues of the temporal context to enhance current representations of the target agent; ii) it aggregates perceptually critical spatial information from heterogeneous agents and overcomes localization errors via multi-scale feature interactions; iii) it integrates multi-source representations of the target agent based on their complementary contributions by an adaptive fusion paradigm. To thoroughly evaluate SCOPE, we consider both real-world and simulated scenarios of collaborative 3D object detection tasks on three datasets. Extensive experiments demonstrate the superiority of our approach and the necessity of the proposed components.

Yang Liu, Jing Liu, Mengyang Zhao et al.

Video anomaly detection is a challenging task in the computer vision community. Most single task-based methods do not consider the independence of unique spatial and temporal patterns, while two-stream structures lack the exploration of the correlations. In this paper, we propose spatial-temporal memories augmented two-stream auto-encoder framework, which learns the appearance normality and motion normality independently and explores the correlations via adversarial learning. Specifically, we first design two proxy tasks to train the two-stream structure to extract appearance and motion features in isolation. Then, the prototypical features are recorded in the corresponding spatial and temporal memory pools. Finally, the encoding-decoding network performs adversarial learning with the discriminator to explore the correlations between spatial and temporal patterns. Experimental results show that our framework outperforms the state-of-the-art methods, achieving AUCs of 98.1% and 89.8% on UCSD Ped2 and CUHK Avenue datasets.

Kun Yang, Jing Liu, Dingkang Yang et al.

With the rapid development of intelligent transportation system applications, a tremendous amount of multi-view video data has emerged to enhance vehicle perception. However, performing video analytics efficiently by exploiting the spatial-temporal redundancy from video data remains challenging. Accordingly, we propose a novel traffic-related framework named CEVAS to achieve efficient object detection using multi-view video data. Briefly, a fine-grained input filtering policy is introduced to produce a reasonable region of interest from the captured images. Also, we design a sharing object manager to manage the information of objects with spatial redundancy and share their results with other vehicles. We further derive a content-aware model selection policy to select detection methods adaptively. Experimental results show that our framework significantly reduces response latency while achieving the same detection accuracy as the state-of-the-art methods.

Jianzhang Zheng, Fan Yang, Hao Shen et al.

Adversarial attacks are often considered as threats to the robustness of Deep Neural Networks (DNNs). Various defending techniques have been developed to mitigate the potential negative impact of adversarial attacks against task predictions. This work analyzes adversarial attacks from a different perspective. Namely, adversarial examples contain implicit information that is useful to the predictions i.e., image classification, and treat the adversarial attacks against DNNs for data self-expression as extracted abstract representations that are capable of facilitating specific learning tasks. We propose an algorithmic framework that leverages the advantages of the DNNs for data self-expression and task-specific predictions, to improve image classification. The framework jointly learns a DNN for attacking Variational Autoencoder (VAE) networks and a DNN for classification, coined as Attacking VAE for Improve Classification (AVIC). The experiment results show that AVIC can achieve higher accuracy on standard datasets compared to the training with clean examples and the traditional adversarial training.

Liang Song, Guangming Wang, Jiuming Liu et al.

In recent studies, the generalization of neural radiance fields for novel view synthesis task has been widely explored. However, existing methods are limited to objects and indoor scenes. In this work, we extend the generalization task to outdoor scenes, trained only on object-level datasets. This approach presents two challenges. Firstly, the significant distributional shift between training and testing scenes leads to black artifacts in rendering results. Secondly, viewpoint changes in outdoor scenes cause ghosting or missing regions in rendered images. To address these challenges, we propose a geometric correction module and an appearance correction module based on multi-head attention mechanisms. We normalize rendered depth and combine it with light direction as query in the attention mechanism. Our network effectively corrects varying scene structures and geometric features in outdoor scenes, generalizing well from object-level to unseen outdoor scenes. Additionally, we use appearance correction module to correct appearance features, preventing rendering artifacts like blank borders and ghosting due to viewpoint changes. By combining these modules, our approach successfully tackles the challenges of outdoor scene generalization, producing high-quality rendering results. When evaluated on four datasets (Blender, DTU, LLFF, Spaces), our network outperforms previous methods. Notably, compared to MVSNeRF, our network improves average PSNR from 19.369 to 25.989, SSIM from 0.838 to 0.889, and reduces LPIPS from 0.265 to 0.224 on Spaces outdoor scenes.

Liangyu Teng, Yang Liu, Jing Liu et al.

In recent years, the e-commerce industry has seen a rapid increase in the demand for advanced AI-driven customer service solutions. Traditional cloud-based models face limitations in terms of latency, personalized services, and privacy concerns. Furthermore, end devices often lack the computational resources to deploy large AI models effectively. In this paper, we propose an innovative End-Cloud Collaboration (ECC) framework for advanced AI customer service in e-commerce. This framework integrates the advantages of large cloud models and mid/small-sized end models by deeply exploring the generalization potential of cloud models and effectively utilizing the computing power resources of terminal chips, alleviating the strain on computing resources to some extent. Specifically, the large cloud model acts as a teacher, guiding and promoting the learning of the end model, which significantly reduces the end model's reliance on large-scale, high-quality data and thereby addresses the data bottleneck in traditional end model training, offering a new paradigm for the rapid deployment of industry applications. Additionally, we introduce an online evolutive learning strategy that enables the end model to continuously iterate and upgrade based on guidance from the cloud model and real-time user feedback. This strategy ensures that the model can flexibly adapt to the rapid changes in application scenarios while avoiding the uploading of sensitive information by performing local fine-tuning, achieving the dual goals of privacy protection and personalized service. %We make systematic contributions to the customized model fine-tuning methods in the e-commerce domain. To conclude, we implement in-depth corpus collection (e.g., data organization, cleaning, and preprocessing) and train an ECC-based industry-specific model for e-commerce customer service.

Hanqi Wang, Xiaoguang Zhu, Tao Chen et al.

Deep learning is widely used to decode the electroencephalogram (EEG) signal. However, there are few attempts to specifically investigate how to explain the EEG-based deep learning models. We conduct a review to summarize the existing works explaining the EEG-based deep learning model. Unfortunately, we find that there is no appropriate method to explain them. Based on the characteristic of EEG data, we suggest a context-aware perturbation method to generate a saliency map from the perspective of the raw EEG signal. Moreover, we also justify that the context information can be used to suppress the artifacts in the EEG-based deep learning model. In practice, some users might want a simple version of the explanation, which only indicates a few features as salient points. To this end, we propose an optional area limitation strategy to restrict the highlighted region. To validate our idea and make a comparison with the other methods, we select three representative EEG-based models to implement experiments on the emotional EEG dataset DEAP. The results of the experiments support the advantages of our method.

Bingning Wang, Haizhou Zhao, Huozhi Zhou et al.

The current generation of large language models (LLMs) is typically designed for broad, general-purpose applications, while domain-specific LLMs, especially in vertical fields like medicine, remain relatively scarce. In particular, the development of highly efficient and practical LLMs for the medical domain is challenging due to the complexity of medical knowledge and the limited availability of high-quality data. To bridge this gap, we introduce Baichuan-M1, a series of large language models specifically optimized for medical applications. Unlike traditional approaches that simply continue pretraining on existing models or apply post-training to a general base model, Baichuan-M1 is trained from scratch with a dedicated focus on enhancing medical capabilities. Our model is trained on 20 trillion tokens and incorporates a range of effective training methods that strike a balance between general capabilities and medical expertise. As a result, Baichuan-M1 not only performs strongly across general domains such as mathematics and coding but also excels in specialized medical fields. We have open-sourced Baichuan-M1-14B, a mini version of our model, which can be accessed through the following links.

Jing Liu, Yang Liu, Jieyu Lin et al.

The increasing utilization of surveillance cameras in smart cities, coupled with the surge of online video applications, has heightened concerns regarding public security and privacy protection, which propelled automated Video Anomaly Detection (VAD) into a fundamental research task within the Artificial Intelligence (AI) community. With the advancements in deep learning and edge computing, VAD has made significant progress and advances synergized with emerging applications in smart cities and video internet, which has moved beyond the conventional research scope of algorithm engineering to deployable Networking Systems for VAD (NSVAD), a practical hotspot for intersection exploration in the AI, IoVT, and computing fields. In this article, we delineate the foundational assumptions, learning frameworks, and applicable scenarios of various deep learning-driven VAD routes, offering an exhaustive tutorial for novices in NSVAD. In addition, this article elucidates core concepts by reviewing recent advances and typical solutions and aggregating available research resources accessible at https://github.com/fdjingliu/NSVAD. Lastly, this article projects future development trends and discusses how the integration of AI and computing technologies can address existing research challenges and promote open opportunities, serving as an insightful guide for prospective researchers and engineers.

Jing Liu, Zhenchao Ma, Zepu Wang et al.

Diffusion models (DMs) have emerged as a powerful class of generative AI models, showing remarkable potential in anomaly detection (AD) tasks across various domains, such as cybersecurity, fraud detection, healthcare, and manufacturing. The intersection of these two fields, termed diffusion models for anomaly detection (DMAD), offers promising solutions for identifying deviations in increasingly complex and high-dimensional data. In this survey, we review recent advances in DMAD research. We begin by presenting the fundamental concepts of AD and DMs, followed by a comprehensive analysis of classic DM architectures including DDPMs, DDIMs, and Score SDEs. We further categorize existing DMAD methods into reconstruction-based, density-based, and hybrid approaches, providing detailed examinations of their methodological innovations. We also explore the diverse tasks across different data modalities, encompassing image, time series, video, and multimodal data analysis. Furthermore, we discuss critical challenges and emerging research directions, including computational efficiency, model interpretability, robustness enhancement, edge-cloud collaboration, and integration with large language models. The collection of DMAD research papers and resources is available at https://github.com/fdjingliu/DMAD.

Zhiyuan Ning, Zhangxun Li, Zhengliang Guo et al.

Video Anomaly Detection (VAD) is an essential yet challenging task in signal processing. Since certain anomalies cannot be detected by isolated analysis of either temporal or spatial information, the interaction between these two types of data is considered crucial for VAD. However, current dual-stream architectures either confine this integral interaction to the bottleneck of the autoencoder or introduce anomaly-irrelevant background pixels into the interactive process, hindering the accuracy of VAD. To address these deficiencies, we propose a Multi-scale Spatial-Temporal Interaction Network (MSTI-Net) for VAD. First, to prioritize the detection of moving objects in the scene and harmonize the substantial semantic discrepancies between the two types of data, we propose an Attention-based Spatial-Temporal Fusion Module (ASTFM) as a substitute for the conventional direct fusion. Furthermore, we inject multi-ASTFM-based connections that bridge the appearance and motion streams of the dual-stream network, thus fostering multi-scale spatial-temporal interaction. Finally, to bolster the delineation between normal and abnormal activities, our system records the regular information in a memory module. Experimental results on three benchmark datasets validate the effectiveness of our approach, which achieves AUCs of 96.8%, 87.6%, and 73.9% on the UCSD Ped2, CUHK Avenue, and ShanghaiTech datasets, respectively.

Jingyu Zhang, Yilei Wang, Lang Qian et al.

As a potential application of Vehicle-to-Everything (V2X) communication, multi-agent collaborative perception has achieved significant success in 3D object detection. While these methods have demonstrated impressive results on standard benchmarks, the robustness of such approaches in the face of complex real-world environments requires additional verification. To bridge this gap, we introduce the first comprehensive benchmark designed to evaluate the robustness of collaborative perception methods in the presence of natural corruptions typical of real-world environments. Furthermore, we propose DSRC, a robustness-enhanced collaborative perception method aiming to learn Density-insensitive and Semantic-aware collaborative Representation against Corruptions. DSRC consists of two key designs: i) a semantic-guided sparse-to-dense distillation framework, which constructs multi-view dense objects painted by ground truth bounding boxes to effectively learn density-insensitive and semantic-aware collaborative representation; ii) a feature-to-point cloud reconstruction approach to better fuse critical collaborative representation across agents. To thoroughly evaluate DSRC, we conduct extensive experiments on real-world and simulated datasets. The results demonstrate that our method outperforms SOTA collaborative perception methods in both clean and corrupted conditions. Code is available at https://github.com/Terry9a/DSRC.

Jing Liu, Zhenchao Ma, Han Yu et al.

Recent advances in collaborative knowledge distillation have demonstrated cutting-edge performance for resource-constrained distributed multimedia learning scenarios. However, achieving such competitiveness requires addressing a fundamental mismatch: high-dimensional teacher knowledge complexity versus heterogeneous client learning capacities, which currently prohibits deployment in edge-based visual analytics systems. Drawing inspiration from curriculum learning principles, we introduce Federated Adaptive Progressive Distillation (FAPD), a consensus-driven framework that orchestrates adaptive knowledge transfer. FAPD hierarchically decomposes teacher features via PCA-based structuring, extracting principal components ordered by variance contribution to establish a natural visual knowledge hierarchy. Clients progressively receive knowledge of increasing complexity through dimension-adaptive projection matrices. Meanwhile, the server monitors network-wide learning stability by tracking global accuracy fluctuations across a temporal consensus window, advancing curriculum dimensionality only when collective consensus emerges. Consequently, FAPD provably adapts knowledge transfer pace while achieving superior convergence over fixed-complexity approaches. Extensive experiments on three datasets validate FAPD's effectiveness: it attains 3.64% accuracy improvement over FedAvg on CIFAR-10, demonstrates 2x faster convergence, and maintains robust performance under extreme data heterogeneity (α=0.1), outperforming baselines by over 4.5%.

Yue Wen, Liang Song, Hesheng Wang

Low-light 3D reconstruction from sparse views remains challenging due to exposure imbalance and degraded color fidelity. While existing methods struggle with view inconsistency and require per-scene training, we propose SplatBright, which is, to our knowledge, the first generalizable 3D Gaussian framework for joint low-light enhancement and reconstruction from sparse sRGB inputs. Our key idea is to integrate physically guided illumination modeling with geometry-appearance decoupling for consistent low-light reconstruction. Specifically, we adopt a dual-branch predictor that provides stable geometric initialization of 3D Gaussian parameters. On the appearance side, illumination consistency leverages frequency priors to enable controllable and cross-view coherent lighting, while an appearance refinement module further separates illumination, material, and view-dependent cues to recover fine texture. To tackle the lack of large-scale geometrically consistent paired data, we synthesize dark views via a physics-based camera model for training. Extensive experiments on public and self-collected datasets demonstrate that SplatBright achieves superior novel view synthesis, cross-view consistency, and better generalization to unseen low-light scenes compared with both 2D and 3D methods.

Juncen Guo, Xiaoguang Zhu, Jingyi Wu et al.

Embodied perception systems face severe challenges of dynamic environment distribution drift when they continuously interact in open physical spaces. However, the existing domain incremental awareness methods often rely on the domain id obtained in advance during the testing phase, which limits their practicability in unknown interaction scenarios. At the same time, the model often overfits to the context-specific perceptual noise, which leads to insufficient generalization ability and catastrophic forgetting. To address these limitations, we propose a domain-id and exemplar-free incremental learning framework for embodied multimedia systems, which aims to achieve robust continuous environment adaptation. This method designs a disentangled representation mechanism to remove non-essential environmental style interference, and guide the model to focus on extracting semantic intrinsic features shared across scenes, thereby eliminating perceptual uncertainty and improving generalization. We further use the weight fusion strategy to dynamically integrate the old and new environment knowledge in the parameter space, so as to ensure that the model adapts to the new distribution without storing historical data and maximally retains the discrimination ability of the old environment. Extensive experiments on multiple standard benchmark datasets show that the proposed method significantly reduces catastrophic forgetting in a completely exemplar-free and domain-id free setting, and its accuracy is better than the existing state-of-the-art methods.

Jiajun Shi, Jian Yang, Jiaheng Liu et al.

Recent advancements in large language models (LLMs) underscore the need for more comprehensive evaluation methods to accurately assess their reasoning capabilities. Existing benchmarks are often domain-specific and thus cannot fully capture an LLM's general reasoning potential. To address this limitation, we introduce the Knowledge Orthogonal Reasoning Gymnasium (KORGym), a dynamic evaluation platform inspired by KOR-Bench and Gymnasium. KORGym offers over fifty games in either textual or visual formats and supports interactive, multi-turn assessments with reinforcement learning scenarios. Using KORGym, we conduct extensive experiments on 19 LLMs and 8 VLMs, revealing consistent reasoning patterns within model families and demonstrating the superior performance of closed-source models. Further analysis examines the effects of modality, reasoning strategies, reinforcement learning techniques, and response length on model performance. We expect KORGym to become a valuable resource for advancing LLM reasoning research and developing evaluation methodologies suited to complex, interactive environments.

Robert Hankache, Kingsley Nketia Acheampong, Liang Song et al.

As large language models (LLMs) are increasingly deployed in multi-turn dialogue and other sustained interactive scenarios, it is essential to understand how extended context affects their performance. Popular benchmarks, focusing primarily on single-turn question answering (QA) tasks, fail to capture the effects of multi-turn exchanges. To address this gap, we introduce a novel set of benchmarks that systematically vary the volume and nature of prior context. We evaluate multiple conventional LLMs, including GPT, Claude, and Gemini, across these benchmarks to measure their sensitivity to contextual variations. Our findings reveal that LLM performance on multiple-choice questions can degrade dramatically in multi-turn interactions, with performance drops as large as 73% for certain models. Even highly capable models such as GPT-4o exhibit up to a 32% decrease in accuracy. Notably, the relative performance of larger versus smaller models is not always predictable. Moreover, the strategic placement of the task description within the context can substantially mitigate performance drops, improving the accuracy by as much as a factor of 3.5. These findings underscore the need for robust strategies to design, evaluate, and mitigate context-related sensitivity in LLMs.

Juncen Guo, Xiaoguang Zhu, Liangyu Teng et al.

Class-incremental Learning (CIL) enables the model to incrementally absorb knowledge from new classes and build a generic classifier across all previously encountered classes. When the model optimizes with new classes, the knowledge of previous classes is inevitably erased, leading to catastrophic forgetting. Addressing this challenge requires making a trade-off between retaining old knowledge and accommodating new information. However, this balancing process often requires sacrificing some information, which can lead to a partial loss in the model's ability to discriminate between classes. To tackle this issue, we design the adaptive weighted parameter fusion with Contrastive Language-Image Pre-training (CLIP), which not only takes into account the variability of the data distribution of different tasks, but also retains all the effective information of the parameter matrix to the greatest extent. In addition, we introduce a balance factor that can balance the data distribution alignment and distinguishability of adjacent tasks. Experimental results on several traditional benchmarks validate the superiority of the proposed method.

Yue Wen, Liang Song, Yijia Liu et al.

Dynamic scene reconstruction for autonomous driving enables vehicles to perceive and interpret complex scene changes more precisely. Dynamic Neural Radiance Fields (NeRFs) have recently shown promising capability in scene modeling. However, many existing methods rely heavily on accurate poses inputs and multi-sensor data, leading to increased system complexity. To address this, we propose FreeDriveRF, which reconstructs dynamic driving scenes using only sequential RGB images without requiring poses inputs. We innovatively decouple dynamic and static parts at the early sampling level using semantic supervision, mitigating image blurring and artifacts. To overcome the challenges posed by object motion and occlusion in monocular camera, we introduce a warped ray-guided dynamic object rendering consistency loss, utilizing optical flow to better constrain the dynamic modeling process. Additionally, we incorporate estimated dynamic flow to constrain the pose optimization process, improving the stability and accuracy of unbounded scene reconstruction. Extensive experiments conducted on the KITTI and Waymo datasets demonstrate the superior performance of our method in dynamic scene modeling for autonomous driving.

Yang Liu, Hongjin Wang, Zepu Wang et al.

Video Anomaly Detection (VAD) remains a fundamental yet formidable task in the video understanding community, with promising applications in areas such as information forensics and public safety protection. Due to the rarity and diversity of anomalies, existing methods only use easily collected regular events to model the inherent normality of normal spatial-temporal patterns in an unsupervised manner. Previous studies have shown that existing unsupervised VAD models are incapable of label-independent data offsets (e.g., scene changes) in real-world scenarios and may fail to respond to light anomalies due to the overgeneralization of deep neural networks. Inspired by causality learning, we argue that there exist causal factors that can adequately generalize the prototypical patterns of regular events and present significant deviations when anomalous instances occur. In this regard, we propose Causal Representation Consistency Learning (CRCL) to implicitly mine potential scene-robust causal variable in unsupervised video normality learning. Specifically, building on the structural causal models, we propose scene-debiasing learning and causality-inspired normality learning to strip away entangled scene bias in deep representations and learn causal video normality, respectively. Extensive experiments on benchmarks validate the superiority of our method over conventional deep representation learning. Moreover, ablation studies and extension validation show that the CRCL can cope with label-independent biases in multi-scene settings and maintain stable performance with only limited training data available.

Juncen Guo, Siao Liu, Xiaoguang Zhu et al.

With the proliferation of multi-modal data in large-scale visual recognition systems, enabling models to continuously acquire knowledge from evolving data streams while preserving prior information has become increasingly critical. Class-Continual Learning (CCL) addresses this challenge by incrementally incorporating new class knowledge without revisiting historical data, making it essential for real-world big data applications. While traditional CCL methods rely solely on visual features, recent advances in Vision-Language Models (VLMs) such as CLIP demonstrate significant potential for CCL by leveraging pre-trained multi-modal knowledge. However, existing approaches face challenges in mitigating catastrophic forgetting while maintaining the cross-modal generalization capabilities of VLMs. To address these limitations, we propose CalFuse, a framework that synergizes feature Calibration with parameter Fusion to enable effective multi-modal knowledge integration in continual learning scenarios. CalFuse introduces a dynamic feature calibration mechanism that adaptively balances original CLIP visual representations with task-specific features, preserving the model's intrinsic cross-modal generalization while adapting to new classes. Concurrently, a QR decomposition-based parameter fusion strategy progressively integrates newly acquired knowledge with historical task parameters, maintaining equilibrium between learning new class representations and retaining prior knowledge across sequential tasks. Extensive experiments on benchmark datasets validate the effectiveness of our approach in large-scale multi-modal continual learning settings, demonstrating superior performance over state-of-the-art methods in both average accuracy and final task retention.

Yuyan Zhou, Liang Song, Bingning Wang et al.

The advent of large language models (LLMs) like GPT-4 has catalyzed the exploration of multi-task learning (MTL), in which a single model demonstrates proficiency across diverse tasks. Task arithmetic has emerged as a cost-effective approach for MTL. It enables performance enhancement across multiple tasks by adding their corresponding task vectors to a pre-trained model. However, the current lack of a method that can simultaneously achieve optimal performance, computational efficiency, and data privacy limits their application to LLMs. In this paper, we propose \textbf{M}odel \textbf{E}xclusive \textbf{T}ask \textbf{A}rithmetic for merging \textbf{GPT}-scale models, which formalizes the objective of model merging into a multi-task learning framework, aiming to minimize the average loss difference between the merged model and each individual task model. Since data privacy limits the use of multi-task training data, we leverage LLMs' local linearity and task vectors' orthogonality to separate the data term and scaling coefficients term and derive a model-exclusive task arithmetic method. Our proposed MetaGPT is data-agnostic and bypasses the heavy search process, making it cost-effective and easy to implement for LLMs.Extensive experiments demonstrate that MetaGPT leads to improvements in task arithmetic and achieves state-of-the-art performance on multiple tasks.

Jinlu Liu, Liang Song, Yongqiang Qin

Few-shot segmentation is a challenging task that aims to segment objects of new classes given scarce support images. In the inductive setting, existing prototype-based methods focus on extracting prototypes from the support images; however, they fail to utilize semantic information of the query images. In this paper, we propose Bi-level Optimization (BiOpt), which succeeds to compute class prototypes from the query images under inductive setting. The learning procedure of BiOpt is decomposed into two nested loops: inner and outer loop. On each task, the inner loop aims to learn optimized prototypes from the query images. An init step is conducted to fully exploit knowledge from both support and query features, so as to give reasonable initialized prototypes into the inner loop. The outer loop aims to learn a discriminative embedding space across different tasks. Extensive experiments on two benchmarks verify the superiority of our proposed BiOpt algorithm. In particular, we consistently achieve the state-of-the-art performance on 5-shot PASCAL-$5^i$ and 1-shot COCO-$20^i$.

Liang Song, Jinlu Liu, Yongqiang Qin

Many Few-Shot Learning research works have two stages: pre-training base model and adapting to novel model. In this paper, we propose to use closed-form base learner, which constrains the adapting stage with pre-trained base model to get better generalized novel model. Following theoretical analysis proves its rationality as well as indication of how to train a well-generalized base model. We then conduct experiments on four benchmarks and achieve state-of-the-art performance in all cases. Notably, we achieve the accuracy of 87.75% on 5-shot miniImageNet which approximately outperforms existing methods by 10%.

Jinlu Liu, Liang Song, Yongqiang Qin

Few-shot learning requires to recognize novel classes with scarce labeled data. Prototypical network is useful in existing researches, however, training on narrow-size distribution of scarce data usually tends to get biased prototypes. In this paper, we figure out two key influencing factors of the process: the intra-class bias and the cross-class bias. We then propose a simple yet effective approach for prototype rectification in transductive setting. The approach utilizes label propagation to diminish the intra-class bias and feature shifting to diminish the cross-class bias. We also conduct theoretical analysis to derive its rationality as well as the lower bound of the performance. Effectiveness is shown on three few-shot benchmarks. Notably, our approach achieves state-of-the-art performance on both miniImageNet (70.31% on 1-shot and 81.89% on 5-shot) and tieredImageNet (78.74% on 1-shot and 86.92% on 5-shot).