Visual emotion recognition (VER), which aims to understand human emotional reactions to different visual stimuli, has garnered increasing attention. However, the inherent ambiguity of emotional features presents significant challenges for data annotation in supervised learning paradigms. To address this limitation, emotion domain adaptation (EDA) facilitates knowledge transfer from labeled source domains to unlabeled target domains. Recently, large visual-language models such as CLIP have demonstrated impressive transfer performance on traditional UDA tasks. However, when generalizing to more abstract concepts such as emotion, the misalignment between CLIP and emotion spaces greatly affects the model performance. To address these challenges, we propose a CLIP-based emotion disentanglement (EmoD) framework designed for EDA. Leveraging perspectives from information bottleneck theory, EmoD implements a disentangler network that extracts emotion-specific features while removing redundant emotion-agnostic information. It also incorporates cross-domain feature alignment to reduce the affective gap between domains. Experimental evaluations in six EDA settings demonstrate that EmoD achieves state-of-the-art performance, surpassing traditional CLIP-based UDA methods by an average of 2.53%.
论文检索
输入标题、作者或关键词，从 801 篇学术成果中精准定位
The evolution of prompt learning methodologies has driven exploration of deeper prompt designs to enhance model performance. However, current deep text prompting approaches suffer from two critical limitations: Over-reliance on constrastive learning objectives that prioritize high-level semantic alignment, neglecting fine-grained feature optimization; Static prompts across all input categories, preventing content-aware adaptation. To address these limitations, we propose AttriPrompt-a novel framework that enhances and refines textual semantic representations by leveraging the intermediate-layer features of CLIP's vision encoder. We designed an Attribute Retrieval module that first clusters visual features from each layer. The aggregated visual features retrieve semantically similar prompts from a prompt pool, which are then concatenated to the input of every layer in the text encoder. Leveraging hierarchical visual information embedded in prompted text features, we introduce Dual-stream Contrastive Learning to realize fine-grained alignment. Furthermore, we introduce a Self-Regularization mechanism by applying explicit regularization constraints between the prompted and non-prompted text features to prevent overfitting on limited training data. Extensive experiments across three benchmarks demonstrate AttriPrompt's superiority over state-of-the-art methods, achieving up to 7.37% improvement in the base-to-novel setting. The observed strength of our method in cross-domain knowledge transfer positions vision-language pre-trained models as more viable solutions for real-world implementation.
Face deepfake detection is a critical technology for verifying the authenticity of facial media content and has long been a focal point in multimedia forensics. However, existing methods face significant challenges, primarily due to their limited ability to generalize across domains. Consequently, the growing variety of forgery techniques, combined with the degradation of visual quality in forged images, makes reliable detection even more difficult. To address these challenges, we propose WKD, a proactive deepfake detection framework based on Watermarking and Knowledge Distillation. The key insights of WKD are twofold: First, we embed watermark information into the Fractional-order Quaternion Radial Harmonic Fourier Moments (FrQRHFMs) space of the host image, achieving a robust balance between imperceptibility and robustness. Second, we design a dual-task learning framework consisting of a watermark extractor and a forgery discriminator, where learnable Low-Rank Adaptation (LoRA) layers are used to transfer knowledge from the extractor to the discriminator, thereby providing additional clues for deepfake detection. Specifically, the integrity of the watermark is compromised only when the host image undergoes a deepfake forgery, while it remains unaffected by conventional attacks. Experimental results on benchmark datasets demonstrate that WKD achieves state-of-the-art performance in both intra-domain and cross-domain deepfake detection, particularly when images are subjected to various conventional attacks.
Recent advances in face forgery detection have shown strong in-domain performance but often fail to generalize to out-of-distribution data, especially when confronted with unseen manipulation techniques or domain shifts (e.g., lighting conditions, camera noise). We propose a novel Mixture-of-Experts framework, termed GM-DF, that decouples domain-specific and domain-invariant features to tackle cross-domain face forgery detection. Our method builds upon a foundation model (CLIP) and incorporates three key modules: (1) Dataset-Embedding Generator that leverages lightweight expert layers and database-aware feature normalization to adaptively modulate features at a per-domain level, capturing idiosyncratic cues without overfitting; (2) Multi-Dataset Representation mechanism that fuses these expert embeddings using scaled dot-product attention and integrates a mask image modeling (MIM) task to amplify local forgery artifacts; (3) Meta-Domain-Embedding Optimizer, inspired by MAML, which alternates between domain-specific (inner-loop) and domain-invariant (outer-loop) updates to facilitate rapid adaptation on new domains. Additionally, inspired by [13] (Yossi Gandelsman, Alexei A Efros, and Jacob Steinhardt. 2024. Interpreting the second-order effects of neurons in clip. arXiv preprint arXiv:2406.04341 (2024)) we introduce second-order feature propagation in the intermediate layers of CLIP to enhance fine-grained artifact cues and propose domain-class disentangled prompts to flexibly encode multi-domain text representations. Together, these strategies enable GM-DF to learn robust, shared forgery cues while preserving essential domain nuances. Our extensive experiments on multiple cross-domain benchmarks demonstrate that GM-DF significantly outperforms state-of-the-art approaches in both detection accuracy and domain transferability, reducing reliance on superficial artifacts and improving generalization to unseen forgeries. Importantly, our design requires minimal overhead beyond standard CLIP, making GM-DF both effective and computationally efficient for real-world face forgery detection.
The rapid evolution of vision-language models (VLMs), such as CLIP, has demonstrated remarkable zero-shot capabilities in downstream tasks. Prompt learning paradigms like Context Optimization (CoOp) refine learnable prompts for efficient adaptation. However, their application in the biomedical domain remains limited due to the insufficient utilization of specialized biomedical knowledge and cross-modality structural relationships. To address these, we introduce MeDKCoOp, a Medical Dual Knowledge-guided graph adaptation method that leverages systematic integration of knowledge through three aspects: exploit domain-specific knowledge from both textual and visual branches, formalize it into graph-structured representations, and leverage knowledge-guided relation transfer for learning cross-modality fusion. By dynamically optimizing learnable prompts through relation learning process, our method achieves disentangled visual representation and enhances transferability to downstream tasks. Evaluations across 8 biomedical datasets spanning 7 imaging modalities demonstrate state-of-the-art cross-domain generalization, with an average 15.12% accuracy improvement over baselines. Our work establishes a new paradigm through graph prompt learning in medical vision-language models, advancing robust diagnostic AI in data-scarce clinical scenarios. Our code is available at: https://github.com/WangYijun-OUC/MeDKCoOp.
Unsupervised domain adaptation transfers knowledge from a labeled source domain to an unlabeled target domain, which assumes that the source domain is labeled correctly, and both source and target data are available during the adaptive process. But collecting large-scale datasets with fully precise annotations is expensive and time-consuming. Besides, due to data privacy and security issues, the source data is often inaccessible during domain adaptation and only unlabeled target data is available. Therefore, considering both the source domain data with noisy labels and the unavailability of source domain data during domain adaptation, this paper proposes an Adaptive Neighbors and Uncertainty Estimation (ANUE) method for Source-Free Unsupervised Domain Adaptation with Noisy Labels (SF-UDA-NL). To the best of our knowledge, there has been no prior research conducted on this particular issue. Specifically, since source domain data contains noisy labels, we propose a reweighted small-loss with uncertainty estimation to filter reliable samples for updating the dual-branch network. To prevent noise knowledge from misleading domain adaptation, we adopt a contrastive learning framework and design an adaptive neighbors module to help target samples generate more reliable pseudo-labels. Then, we design a reweighted contrastive loss at both class level and instance level based on uncertainty estimation to further enhance the network's classification performance. We conduct extensive experiments on three widely used datasets for unsupervised domain adaptation in image classification, and the results demonstrate the effectiveness and robustness of our method.
Learning from large-scale pre-trained models with strong generalization ability has shown remarkable success in a wide range of downstream tasks recently, but it is still underexplored in the challenging few-shot class-incremental learning (FSCIL) task. It aims to continually learn new concepts from limited training samples without forgetting the old ones at the same time. In this paper, we introduce DSS-Prompt, a simple yet effective approach that transforms the pre-trained Vision Transformer with minimal modifications in the way of prompts into a strong FSCIL classifier. Concretely, we synergistically utilize two complementary types of prompts in each Transformer block: static prompts to bridge the domain gap between the pre-training and downstream datasets, thus enabling better adaption; and dynamic prompts to capture instance-aware semantics, thus enabling easy transfer from base to novel classes. Specially, to generate dynamic prompts, we leverage a pre-trained multi-modal model to extract input-related diverse semantics, thereby generating complementary input-aware prompts, and then adaptively adjust their importance across different layers. In this way, on top of the prompted visual embeddings, a simple prototype classifier can beat state-of-the-arts without further training on the incremental tasks. We conduct extensive experiments on four benchmarks to validate the effectiveness of our DSS-Prompt and show that it consistently achieves better performance than existing approaches on all datasets and can alleviate the catastrophic forgetting issue as well.
Lifelong Person Re-identification (LReID) focuses on continuously adapting to new domains over time while preserving knowledge from previously seen domains, particularly under the domain incremental learning setting. The major challenge of LReID is catastrophic forgetting, typically caused by large domain shifts during training. To address this, we propose a novel Amplitude-aware Domain Style Replay (ADSR) framework, which introduces a Fourier-based Style Transfer (FST) mechanism to generate synthetic data that reflects the style of previously encountered domains. These proxy images help retain prior knowledge without the need to store actual past data. Our method transfers stylistic information-mainly encoded in the amplitude spectrum-from old domains to new ones, creating old-stylized images that preserve the content of new domain data while adopting the visual style of earlier domains. To further boost generalization, we design a Self-Stylization Normalization (SSN) module that adapts the current domain's style distribution, making the model more robust to stylistic variations. Additionally, we introduce a Multi-Granularity Transfer (MGT) module that uses K-Means clustering to extract multiple representative style features from each domain, enabling compact yet comprehensive storage and replay of domain-specific information. Extensive experiments on multiple LReID benchmarks show that ADSR achieves superior performance over existing approaches, effectively reducing forgetting and improving cross-domain generalization. Our code is available at https://github.com/cclong8/MM2025-ADSR.
Adversarial patches are widely used to evaluate the robustness of object detection systems in real-world scenarios. These patches were initially designed to deceive single-modal detectors (e.g., visible or infrared) and have recently been extended to target visible-infrared dual-modal detectors. However, existing dual-modal adversarial patch attacks have limited attack effectiveness across diverse physical scenarios. To address this, we propose CDUPatch, a universal cross-modal patch attack against visible-infrared object detectors across scales, views, and scenarios. Specifically, we observe that color variations lead to different levels of thermal absorption, resulting in temperature differences in infrared imaging. Leveraging this property, we propose an RGB-to-infrared adapter that maps RGB patches to infrared patches, enabling unified optimization of cross-modal patches. By learning an optimal color distribution on the adversarial patch, we can manipulate its thermal response and generate an adversarial infrared texture. Additionally, we introduce a multi-scale clipping strategy and construct a new visible-infrared dataset, MSDrone, which contains aerial vehicle images in varying scales and perspectives. These data augmentation strategies enhance the robustness of our patch in real-world conditions. Experiments on four benchmark datasets (e.g., DroneVehicle, LLVIP, VisDrone, MSDrone) show that our method outperforms existing patch attacks in the digital domain. Extensive physical tests further confirm strong transferability across scales, views, and scenarios. Attack demos are provided in the supplementary materials.
Knowledge Distillation (KD) compresses neural networks by learning a small network (student) via transferring knowledge from a pre-trained large network (teacher). Many endeavours have been devoted to the image domain, while few works focus on video analysis which desires training much larger model making it be hardly deployed in resource-limited devices. However, traditional methods neglect two important problems, i.e., 1) Since the capacity gap between the teacher and the student exists, some knowledge w.r.t. difficult-to-transfer samples cannot be correctly transferred, or even badly affects the final performance of student, and 2) As training progresses, difficult-to-transfer samples may become easier to learn, and vice versa. To alleviate the two problems, we propose a Sample-level Adaptive Knowledge Distillation (SAKD) framework for action recognition. In particular, it mainly consists of the sample distillation difficulty evaluation module and the sample adaptive distillation module. The former applies the temporal interruption to frames, i.e., randomly dropout or shuffle the frames during training, which increases the learning difficulty of samples during distillation, so as to better discriminate their distillation difficulty. The latter module adaptively adjusts distillation ratio at sample level, such that KD loss dominates the training with easy-to-transfer samples while vanilla loss dominates that with difficult-to-transfer samples. More importantly, we only select those samples with both low distillation difficulty and high diversity to train the student model for reducing computational cost. Experimental results on three video benchmarks and one image benchmark demonstrate the superiority of the proposed method by striking a good balance between performance and efficiency. Code is available at https://github.com/mlvccn/SAKD_ActionRec.
Document Image Tampering Localization (DITL) advances considerably, yet achieving robust cross-dataset generalization remains a formidable challenge for practical applications. Expanding existing document datasets for training is labor-intensive, making it appealing to incorporate data from non-document domains such as natural scene images. However, domain-specific variations, including differences in color distribution and texture, compromise the performance of joint training. To address this issue, we propose DITL2, a Dual-stage Invariance Transfer Learning framework for Document Image Tampering Localization that consists of Cross-Domain Invariance Pre-training (CDIP) and Frequency Decoupling Parameter Adaptation (FDPA). In the pre-training stage, CDIP employs style transfer and texture consistency learning to suppress domain-specific influences from tampered natural scene images, and tampering trace commonality learning to acquire domain-invariant features. In the fine-tuning stage, FDPA adapts the parameters of the pre-trained model, leveraging the general knowledge from the pre-trained model to address DITL tasks while reducing the risk of overfitting. Experiments show that this approach effectively leverages external data resources to boost model performance, achieving state-of-the-art results across a variety of cross-dataset settings.
Cross-Domain Sequential Recommendation (CDSR) has recently gained attention for countering data sparsity by transferring knowledge across domains. A common approach merges domain-specific sequences into cross-domain sequences, serving as bridges to connect domains. One key challenge is to correctly extract the shared knowledge among these sequences and appropriately transfer it. Most existing works directly transfer unfiltered cross-domain knowledge rather than extracting domain-invariant components and adaptively integrating them into domain-specific modelings. Another challenge lies in aligning the domain-specific and cross-domain sequences. Existing methods align these sequences based on timestamps, but this approach can cause prediction mismatches when the current tokens and their targets belong to different domains. In such cases, the domain-specific knowledge carried by the current tokens may degrade performance. To address these challenges, we propose the A-B-Cross-to-Invariant Learning Recommender (ABXI). Specifically, leveraging LoRA's effectiveness for efficient adaptation, ABXI incorporates two types of LoRAs to facilitate knowledge adaptation. First, all sequences are processed through a shared encoder that employs a domain LoRA for each sequence, thereby preserving unique domain characteristics. Next, we introduce an invariant projector that extracts domain-invariant interests from cross-domain representations, utilizing an invariant LoRA to adapt these interests into modeling each specific domain. Besides, to avoid prediction mismatches, all domain-specific sequences are aligned to match the domains of the cross-domain ground truths. Experimental results on three datasets demonstrate that our approach outperforms other CDSR counterparts by a large margin. The codes are available in https://github.com/DiMarzioBian/ABXI.
Graph Neural Networks (GNNs) have achieved remarkable success in node classification tasks on individual graphs. However, existing GNNs trained within a specific domain ( a.k.a., source domain) frequently exhibit unsatisfied performance when transferred to another domain ( a.k.a., target domain), due to the domain gap. To tackle this issue, Few Shot Graph Domain Adaptation (FSGDA) is introduced to the node classification task, facilitating knowledge transfer from a fully labeled source graph to a target graph with minimal annotations for each class. An intuitive solution is directly training the GNN with labeled source and target samples together. Nevertheless, there are two issues in this procedure: (1) When the annotations on the target domain used for training are extremely sparse, the GNN performance may significantly be damaged by nodes with the source-domain bias not aligning with the target-domain distribution. (2) Apart from the biased nodes, the low-value nodes among the remaining nodes impede the GNN learning for the core nodes, like the limited target training nodes. To address the above issues, we propose a new method for FSGDA, named GraphInflu, whose core idea is to grasp the key takeaways from the source domain to facilitate the adaptation process. It contains two characteristic modules, including the Supportive Node Selector and the Soft Logic-Inspired Node Reweighting. The former aims to identify the most influential set of source nodes based on their contribution to improving performance on target nodes. The latter further focuses more on the core nodes in the selected influential set, which closely align with the target nodes especially those presenting challenging predictions. Extensive experiments validate the efficacy of GraphInflu by overcoming the current state-of-the-art methods. Our code is available at https://github.com/lvXiangwei/GraphInflu.git.
ROMA: Recommendation-Oriented Language Model Adaptation Using Multi-Modal Multi-Domain Item Sequences
PDF ↗Sequential recommendation (SR) aims to capture dynamic user preferences from users' historical behaviors. Recently, benefiting from astonishing understanding ability of pre-trained language models (PLMs), text-enhanced sequential recommender becomes a promising direction, which employs PLMs to extract semantic information for user/item representation. Although promising in improving performance and transferability, few existing text-enhanced SR studies have analyzed the differences between PLMs and recommenders, restricting the ability of PLMs for recommendation. In this paper, we make an in-depth comparison and conclude their discrepancies in representation and knowledge level, respectively, caused by different multi-modal content and task-oriented capabilities. Based on this, we propose a Recommendation-Oriented Language Model Adaptation framework (named ROMA) using multi-modal multi-domain item sequences. To empower PLMs with a rational understanding of user/item modeling and the recommendation task, ROMA partitions a PLM into bottom and top layers, respectively, allowing representation-level and task-level adaptation with elaborately designed architectures, transferring strategy and learning framework. Our experimental results on public benchmarks demonstrate the effectiveness and transferability of our framework. Additionally, we showcase the application value of ROMA on the recommender system of Huawei's AppGallery through online A/B testing, which shows significant improvements in online metrics.
A long-standing challenge for pushing sensor-based human activity recognition (HAR) to industrial usage is the distribution shift between training data and testing data: significant variations in data distribution lead to a notable decline in performance. Recently, Large Language Models (LLMs) have demonstrated exceptional generalization capability, which provides a new opportunity to mitigate the distribution shift problem of HAR. However, since LLMs are inherently designed and trained on textual data, their potential to enhance generalization in HAR applications remains an open question. In this paper, we introduce LLM4HAR, a novel LLM-based model to improve cross-domain HAR. LLM4HAR consists of three main modules: (i) the Sensor Data Adaptation module, which aligns IMU signals with LLMs via sensor embedding(ii) the Sensor Knowledge Learning module, which injects sensor knowledge into LLMs for activity recognition, and (iii) the Efficiency Enhancement module, which employs a partial training strategy and reduces the model size by more than 10 times. Extensive evaluations show that LLM4HAR outperforms the existing methods by 13.82% in average F1 score, demonstrating the feasibility and effectiveness of transferring knowledge from pretrained LLMs to enhance HAR. Further, LLM4HAR has been adopted by JD Logistics to support downstream applications such as Courier Welfare Improvement and Map Data Generation.
Cross-domain recommendation (CDR) has gained significant attention for its ability to address data sparsity issue. However, most existing CDR methods focus primarily on improving recommendation accuracy while largely overlooking fairness considerations, which can lead to biased outcomes and unfair treatment of different user groups. To solve this critical problem, we investigate whether fairness can be transferred from the source domain to the target domain. Our analysis suggests that fairness can be effectively transferred if the fairness of the source domain is ensured and the distributions of the source and target domains are well aligned. Based on this, we propose the FairCDR, a novel framework that can achieve the knowledge transfer of fairness and user preferences simultaneously. FairCDR owns two phases: single-domain fairness guarantee and inter-domain distribution alignment. In the first phase, we employ an adversarial learning-based recommender (ALR) to disentangle user preferences from sensitive attributes in the source domain. In the second phase, we introduce a new mutual learning-based diffusion model (MLDiff), which engages in mutual learning with ALR to progressively align the distributions of the source and target domains. This improves ALR's adaptability to distribution shifts, ultimately ensuring fairness and recommendation performance in the target domain. Extensive experiments on multiple real-world cross-domain datasets demonstrate that FairCDR surpasses existing strong baselines in both fairness and recommendation quality.
FreRA: A Frequency-Refined Augmentation for Contrastive Learning on Time Series Classification
PDF ↗Contrastive learning has emerged as a competent approach for unsupervised representation learning. However, the design of an optimal augmentation strategy, although crucial for contrastive learning, is less explored for time series classification tasks. Existing predefined time-domain augmentation methods are primarily adopted from vision and are not specific to time series data. Consequently, this cross-modality incompatibility may distort the semantically relevant information of time series by introducing mismatched patterns into the data. To address this limitation, we present a novel perspective from the frequency domain and identify three advantages for downstream classification: 1) the frequency component naturally encodes global features, 2) the orthogonal nature of the Fourier basis allows easier isolation and independent modifications of critical and unimportant information, and 3) a compact set of frequency components can preserve semantic integrity. To fully utilize the three properties, we propose the lightweight yet effective Frequency-Refined Augmentation (FreRA) tailored for time series contrastive learning on classification tasks, which can be seamlessly integrated with contrastive learning frameworks in a plug-and-play manner. Specifically, FreRA automatically separates critical and unimportant frequency components. Accordingly, we propose semantic-aware Identity Modification and semantic-agnostic Self-adaptive Modification to protect semantically relevant information in the critical frequency components and infuse variance into the unimportant ones respectively. Theoretically, we prove that FreRA generates semantic-preserving views. Empirically, we conduct extensive experiments on two benchmark datasets, including UCR and UEA archives, as well as five large-scale datasets on diverse applications. FreRA consistently outperforms ten leading baselines on time series classification, anomaly detection, and transfer learning tasks, demonstrating superior capabilities in contrastive representation learning and generalization in transfer learning scenarios across diverse datasets. The code is available at https://github.com/Tian0426/FreRA.
Link Prediction (LP) is a critical task in graph machine learning. While Graph Neural Networks (GNNs) have significantly advanced LP performance recently, existing methods face key challenges including limited supervision from sparse connectivity, sensitivity to initialization, and poor generalization under distribution shifts. We explore pretraining as a solution to address these challenges. Unlike node classification, LP is inherently a pairwise task, which requires the integration of both node- and edge-level information. In this work, we present the first systematic study on the transferability of these distinct modules and propose a late fusion strategy to effectively combine their outputs for improved performance. To handle the diversity of pretraining data and avoid negative transfer, we introduce a Mixture-of-Experts (MoE) framework that captures distinct patterns in separate experts, facilitating seamless application of the pretrained model on diverse downstream datasets. For fast adaptation, we develop a parameter-efficient tuning strategy that allows the pretrained model to adapt to unseen datasets with minimal computational overhead. Experiments on 16 datasets across two domains demonstrate the effectiveness of our approach, achieving state-of-the-art performance on low-resource link prediction while obtaining competitive results compared to end-to-end trained methods, with over 10,000x lower computational overhead.
Imputation via Domain Adaptation: Rethinking Variable Subset Forecasting from Knowledge Transfer
PDF ↗Multivariate time series forecasting in practical deployment faces a critical challenge termed Variable Subset Forecasting (VSF), where certain variables accessible during training are entirely missing during inference. This creates a stark discrepancy between the training (source domain with full variables) and inference (target domain with partial variables) environments, disrupting cross-variable dependencies and fragmenting global temporal patterns. Existing imputation methods, limited to transferring local knowledge (e.g., temporal neighbors or pairwise correlations), fail to capture essential global dynamics, leading to severe performance degradation under distribution shifts. To address these challenges, we redefine VSF as a cross-domain knowledge transfer problem and propose VIDA, a framework that systematically transfers Variable Invariant knowledge from complete to partial observations through Domain Adaptation. Key to our approach is (1) Global time-frequency joint representation learning, which encodes temporal dynamics via dilated convolutions and captures low-frequency spectral consistency using Fourier neural operators, and (2) Sinkhorn-regularized distribution alignment to bridge non-overlapping feature supports across domains via optimal transport. Unlike imputation-first methods, VIDA enforces task-driven consistency by jointly optimizing predictions on reconstructed and original data, ensuring the transferred knowledge directly enhances forecasting robustness. Extensive experiments across four real-world datasets show that VIDA outperforms state-of-the-art imputation methods by 25% on average with partially observed variables. This work establishes a new paradigm for variable-missing scenarios by unifying imputation and forecasting through principled knowledge transfer.
Classical time-series forecasting methods rely on learning from a fixed context window of past observations and extrapolating to a fixed number of future steps. They are therefore limited in flexibility and effectiveness in modeling tasks of varying historical lengths or future horizons, requiring re-training for each setting. We propose Contextual Meta-Adaptation (CMA), a unified scheme that captures history contexts and extrapolates future horizons with a single training. CMA integrates denoising diffusion capability which can sequentially generate a future series with maximum likelihood and remains consistent with observed history and past generations. We develop an efficient update strategy using gradient-based meta-learning for full-parameter tuning during training and LoRA adaptation at testing time. We validate our CMA scheme on in-context learning with longer observed history, ex(tend)-context learning with up to 3 times longer future horizons, and transfer-learning across different source-target domains. Our CMA outperforms existing methods by an average of 7% on six academic datasets and a significant 16% improvement on four realistic marketing datasets.