Wenwen Qiang

Xingzhe Su, Wenwen Qiang, Jie Hu et al.

Generative adversarial networks (GANs) have achieved remarkable progress in the natural image field. However, when applying GANs in the remote sensing (RS) image generation task, an extraordinary phenomenon is observed: the GAN model is more sensitive to the size of training data for RS image generation than for natural image generation. In other words, the generation quality of RS images will change significantly with the number of training categories or samples per category. In this paper, we first analyze this phenomenon from two kinds of toy experiments and conclude that the amount of feature information contained in the GAN model decreases with reduced training data. Then we establish a structural causal model (SCM) of the data generation process and interpret the generated data as the counterfactuals. Based on this SCM, we theoretically prove that the quality of generated images is positively correlated with the amount of feature information. This provides insights for enriching the feature information learned by the GAN model during training. Consequently, we propose two innovative adjustment schemes, namely Uniformity Regularization (UR) and Entropy Regularization (ER), to increase the information learned by the GAN model at the distributional and sample levels, respectively. We theoretically and empirically demonstrate the effectiveness and versatility of our methods. Extensive experiments on three RS datasets and two natural datasets show that our methods outperform the well-established models on RS image generation tasks. The source code is available at https://github.com/rootSue/Causal-RSGAN.

Jiangmeng Li, Yanan Zhang, Wenwen Qiang et al.

Few-shot learning models learn representations with limited human annotations, and such a learning paradigm demonstrates practicability in various tasks, e.g., image classification, object detection, etc. However, few-shot object detection methods suffer from an intrinsic defect that the limited training data makes the model cannot sufficiently explore semantic information. To tackle this, we introduce knowledge distillation to the few-shot object detection learning paradigm. We further run a motivating experiment, which demonstrates that in the process of knowledge distillation, the empirical error of the teacher model degenerates the prediction performance of the few-shot object detection model as the student. To understand the reasons behind this phenomenon, we revisit the learning paradigm of knowledge distillation on the few-shot object detection task from the causal theoretic standpoint, and accordingly, develop a Structural Causal Model. Following the theoretical guidance, we propose a backdoor adjustment-based knowledge distillation method for the few-shot object detection task, namely Disentangle and Remerge (D&R), to perform conditional causal intervention toward the corresponding Structural Causal Model. Empirically, the experiments on benchmarks demonstrate that D&R can yield significant performance boosts in few-shot object detection. Code is available at https://github.com/ZYN-1101/DandR.git.

Yujie Zhou, Wenwen Qiang, Anyi Rao et al.

Zero-shot skeleton-based action recognition aims to recognize actions of unseen categories after training on data of seen categories. The key is to build the connection between visual and semantic space from seen to unseen classes. Previous studies have primarily focused on encoding sequences into a singular feature vector, with subsequent mapping the features to an identical anchor point within the embedded space. Their performance is hindered by 1) the ignorance of the global visual/semantic distribution alignment, which results in a limitation to capture the true interdependence between the two spaces. 2) the negligence of temporal information since the frame-wise features with rich action clues are directly pooled into a single feature vector. We propose a new zero-shot skeleton-based action recognition method via mutual information (MI) estimation and maximization. Specifically, 1) we maximize the MI between visual and semantic space for distribution alignment; 2) we leverage the temporal information for estimating the MI by encouraging MI to increase as more frames are observed. Extensive experiments on three large-scale skeleton action datasets confirm the effectiveness of our method. Code: https://github.com/YujieOuO/SMIE.

Dengsheng Chen, Jie Hu, Wenwen Qiang et al.

Over past few years afterward the birth of ResNet, skip connection has become the defacto standard for the design of modern architectures due to its widespread adoption, easy optimization and proven performance. Prior work has explained the effectiveness of the skip connection mechanism from different perspectives. In this work, we deep dive into the model's behaviors with skip connections which can be formulated as a learnable Markov chain. An efficient Markov chain is preferred as it always maps the input data to the target domain in a better way. However, while a model is explained as a Markov chain, it is not guaranteed to be optimized following an efficient Markov chain by existing SGD-based optimizers which are prone to get trapped in local optimal points. In order to towards a more efficient Markov chain, we propose a simple routine of penal connection to make any residual-like model become a learnable Markov chain. Aside from that, the penal connection can also be viewed as a particular model regularization and can be easily implemented with one line of code in the most popular deep learning frameworks~\footnote{Source code: \url{https://github.com/densechen/penal-connection}}. The encouraging experimental results in multi-modal translation and image recognition empirically confirm our conjecture of the learnable Markov chain view and demonstrate the superiority of the proposed penal connection.

Xingyu Zhang, Siyu Zhao, Zeen Song et al.

Long-term time series forecasting is a long-standing challenge in various applications. A central issue in time series forecasting is that methods should expressively capture long-term dependency. Furthermore, time series forecasting methods should be flexible when applied to different scenarios. Although Fourier analysis offers an alternative to effectively capture reusable and periodic patterns to achieve long-term forecasting in different scenarios, existing methods often assume high-frequency components represent noise and should be discarded in time series forecasting. However, we conduct a series of motivation experiments and discover that the role of certain frequencies varies depending on the scenarios. In some scenarios, removing high-frequency components from the original time series can improve the forecasting performance, while in others scenarios, removing them is harmful to forecasting performance. Therefore, it is necessary to treat the frequencies differently according to specific scenarios. To achieve this, we first reformulate the time series forecasting problem as learning a transfer function of each frequency in the Fourier domain. Further, we design Frequency Dynamic Fusion (FreDF), which individually predicts each Fourier component, and dynamically fuses the output of different frequencies. Moreover, we provide a novel insight into the generalization ability of time series forecasting and propose the generalization bound of time series forecasting. Then we prove FreDF has a lower bound, indicating that FreDF has better generalization ability. Extensive experiments conducted on multiple benchmark datasets and ablation studies demonstrate the effectiveness of FreDF. The code is available at https://github.com/Zh-XY22/FreDF.

Jiexin Wang, Yujie Zhou, Wenwen Qiang et al.

Human motion prediction (HMP) has emerged as a popular research topic due to its diverse applications, but it remains a challenging task due to the stochastic and aperiodic nature of future poses. Traditional methods rely on hand-crafted features and machine learning techniques, which often struggle to model the complex dynamics of human motion. Recent deep learning-based methods have achieved success by learning spatio-temporal representations of motion, but these models often overlook the reliability of motion data. Additionally, the temporal and spatial dependencies of skeleton nodes are distinct. The temporal relationship captures motion information over time, while the spatial relationship describes body structure and the relationships between different nodes. In this paper, we propose a novel spatio-temporal branching network using incremental information for HMP, which decouples the learning of temporal-domain and spatial-domain features, extracts more motion information, and achieves complementary cross-domain knowledge learning through knowledge distillation. Our approach effectively reduces noise interference and provides more expressive information for characterizing motion by separately extracting temporal and spatial features. We evaluate our approach on standard HMP benchmarks and outperform state-of-the-art methods in terms of prediction accuracy.

Jiangmeng Li, Wenwen Qiang, Changwen Zheng et al.

While self-supervised learning techniques are often used to mining implicit knowledge from unlabeled data via modeling multiple views, it is unclear how to perform effective representation learning in a complex and inconsistent context. To this end, we propose a methodology, specifically consistency and complementarity network (CoCoNet), which avails of strict global inter-view consistency and local cross-view complementarity preserving regularization to comprehensively learn representations from multiple views. On the global stage, we reckon that the crucial knowledge is implicitly shared among views, and enhancing the encoder to capture such knowledge from data can improve the discriminability of the learned representations. Hence, preserving the global consistency of multiple views ensures the acquisition of common knowledge. CoCoNet aligns the probabilistic distribution of views by utilizing an efficient discrepancy metric measurement based on the generalized sliced Wasserstein distance. Lastly on the local stage, we propose a heuristic complementarity-factor, which joints cross-view discriminative knowledge, and it guides the encoders to learn not only view-wise discriminability but also cross-view complementary information. Theoretically, we provide the information-theoretical-based analyses of our proposed CoCoNet. Empirically, to investigate the improvement gains of our approach, we conduct adequate experimental validations, which demonstrate that CoCoNet outperforms the state-of-the-art self-supervised methods by a significant margin proves that such implicit consistency and complementarity preserving regularization can enhance the discriminability of latent representations.

Jiangmeng Li, Wenyi Mo, Wenwen Qiang et al.

Vision-language models are pre-trained by aligning image-text pairs in a common space to deal with open-set visual concepts. To boost the transferability of the pre-trained models, recent works adopt fixed or learnable prompts, i.e., classification weights are synthesized from natural language describing task-relevant categories, to reduce the gap between tasks in the training and test phases. However, how and what prompts can improve inference performance remains unclear. In this paper, we explicitly clarify the importance of including semantic information in prompts, while existing prompting methods generate prompts without exploring the semantic information of textual labels. Manually constructing prompts with rich semantics requires domain expertise and is extremely time-consuming. To cope with this issue, we propose a semantic-aware prompt learning method, namely CPKP, which retrieves an ontological knowledge graph by treating the textual label as a query to extract task-relevant semantic information. CPKP further introduces a double-tier confounder-pruning procedure to refine the derived semantic information. The graph-tier confounders are gradually identified and phased out, inspired by the principle of Granger causality. The feature-tier confounders are demolished by following the maximum entropy principle in information theory. Empirically, the evaluations demonstrate the effectiveness of CPKP, e.g., with two shots, CPKP outperforms the manual-prompt method by 4.64% and the learnable-prompt method by 1.09% on average, and the superiority of CPKP in domain generalization compared to benchmark approaches. Our implementation is available at https://github.com/Mowenyii/CPKP.

Hang Gao, Jiangmeng Li, Wenwen Qiang et al.

The prevailing graph neural network models have achieved significant progress in graph representation learning. However, in this paper, we uncover an ever-overlooked phenomenon: the pre-trained graph representation learning model tested with full graphs underperforms the model tested with well-pruned graphs. This observation reveals that there exist confounders in graphs, which may interfere with the model learning semantic information, and current graph representation learning methods have not eliminated their influence. To tackle this issue, we propose Robust Causal Graph Representation Learning (RCGRL) to learn robust graph representations against confounding effects. RCGRL introduces an active approach to generate instrumental variables under unconditional moment restrictions, which empowers the graph representation learning model to eliminate confounders, thereby capturing discriminative information that is causally related to downstream predictions. We offer theorems and proofs to guarantee the theoretical effectiveness of the proposed approach. Empirically, we conduct extensive experiments on a synthetic dataset and multiple benchmark datasets. The results demonstrate that compared with state-of-the-art methods, RCGRL achieves better prediction performance and generalization ability.

Jiangmeng Li, Wenwen Qiang, Changwen Zheng et al.

What matters for contrastive learning? We argue that contrastive learning heavily relies on informative features, or "hard" (positive or negative) features. Early works include more informative features by applying complex data augmentations and large batch size or memory bank, and recent works design elaborate sampling approaches to explore informative features. The key challenge toward exploring such features is that the source multi-view data is generated by applying random data augmentations, making it infeasible to always add useful information in the augmented data. Consequently, the informativeness of features learned from such augmented data is limited. In response, we propose to directly augment the features in latent space, thereby learning discriminative representations without a large amount of input data. We perform a meta learning technique to build the augmentation generator that updates its network parameters by considering the performance of the encoder. However, insufficient input data may lead the encoder to learn collapsed features and therefore malfunction the augmentation generator. A new margin-injected regularization is further added in the objective function to avoid the encoder learning a degenerate mapping. To contrast all features in one gradient back-propagation step, we adopt the proposed optimization-driven unified contrastive loss instead of the conventional contrastive loss. Empirically, our method achieves state-of-the-art results on several benchmark datasets.

Wenwen Qiang, Jiangmeng Li, Changwen Zheng et al.

Unsupervised domain adaptation (UDA) requires source domain samples with clean ground truth labels during training. Accurately labeling a large number of source domain samples is time-consuming and laborious. An alternative is to utilize samples with noisy labels for training. However, training with noisy labels can greatly reduce the performance of UDA. In this paper, we address the problem that learning UDA models only with access to noisy labels and propose a novel method called robust local preserving and global aligning network (RLPGA). RLPGA improves the robustness of the label noise from two aspects. One is learning a classifier by a robust informative-theoretic-based loss function. The other is constructing two adjacency weight matrices and two negative weight matrices by the proposed local preserving module to preserve the local topology structures of input data. We conduct theoretical analysis on the robustness of the proposed RLPGA and prove that the robust informative-theoretic-based loss and the local preserving module are beneficial to reduce the empirical risk of the target domain. A series of empirical studies show the effectiveness of our proposed RLPGA.

Wenwen Qiang, Jiangmeng Li, Changwen Zheng et al.

Contrastive learning (CL)-based self-supervised learning models learn visual representations in a pairwise manner. Although the prevailing CL model has achieved great progress, in this paper, we uncover an ever-overlooked phenomenon: When the CL model is trained with full images, the performance tested in full images is better than that in foreground areas; when the CL model is trained with foreground areas, the performance tested in full images is worse than that in foreground areas. This observation reveals that backgrounds in images may interfere with the model learning semantic information and their influence has not been fully eliminated. To tackle this issue, we build a Structural Causal Model (SCM) to model the background as a confounder. We propose a backdoor adjustment-based regularization method, namely Interventional Contrastive Learning with Meta Semantic Regularizer (ICL-MSR), to perform causal intervention towards the proposed SCM. ICL-MSR can be incorporated into any existing CL methods to alleviate background distractions from representation learning. Theoretically, we prove that ICL-MSR achieves a tighter error bound. Empirically, our experiments on multiple benchmark datasets demonstrate that ICL-MSR is able to improve the performances of different state-of-the-art CL methods.

Hang Gao, Jiangmeng Li, Wenwen Qiang et al.

Benefiting from the injection of human prior knowledge, graphs, as derived discrete data, are semantically dense so that models can efficiently learn the semantic information from such data. Accordingly, graph neural networks (GNNs) indeed achieve impressive success in various fields. Revisiting the GNN learning paradigms, we discover that the relationship between human expertise and the knowledge modeled by GNNs still confuses researchers. To this end, we introduce motivating experiments and derive an empirical observation that the GNNs gradually learn human expertise in general domains. By further observing the ramifications of introducing expertise logic into graph representation learning, we conclude that leading the GNNs to learn human expertise can improve the model performance. Hence, we propose a novel graph representation learning method to incorporate human expert knowledge into GNN models. The proposed method ensures that the GNN model can not only acquire the expertise held by human experts but also engage in end-to-end learning from datasets. Plentiful experiments on the crafted and real-world domains support the consistent effectiveness of the proposed method.

Jingyao Wang, Wenwen Qiang, Xingzhe Su et al.

Meta-learning aims to learn general knowledge with diverse training tasks conducted from limited data, and then transfer it to new tasks. It is commonly believed that increasing task diversity will enhance the generalization ability of meta-learning models. However, this paper challenges this view through empirical and theoretical analysis. We obtain three conclusions: (i) there is no universal task sampling strategy that can guarantee the optimal performance of meta-learning models; (ii) over-constraining task diversity may incur the risk of under-fitting or over-fitting during training; and (iii) the generalization performance of meta-learning models are affected by task diversity, task entropy, and task difficulty. Based on this insight, we design a novel task sampler, called Adaptive Sampler (ASr). ASr is a plug-and-play module that can be integrated into any meta-learning framework. It dynamically adjusts task weights according to task diversity, task entropy, and task difficulty, thereby obtaining the optimal probability distribution for meta-training tasks. Finally, we conduct experiments on a series of benchmark datasets across various scenarios, and the results demonstrate that ASr has clear advantages.

Lingyu Si, Hongwei Dong, Wenwen Qiang et al.

Due to limitations in data quality, some essential visual tasks are difficult to perform independently. Introducing previously unavailable information to transfer informative dark knowledge has been a common way to solve such hard tasks. However, research on why transferred knowledge works has not been extensively explored. To address this issue, in this paper, we discover the correlation between feature discriminability and dimensional structure (DS) by analyzing and observing features extracted from simple and hard tasks. On this basis, we express DS using deep channel-wise correlation and intermediate spatial distribution, and propose a novel cross-modal knowledge distillation (CMKD) method for better supervised cross-modal learning (CML) performance. The proposed method enforces output features to be channel-wise independent and intermediate ones to be uniformly distributed, thereby learning semantically irrelevant features from the hard task to boost its accuracy. This is especially useful in specific applications where the performance gap between dual modalities is relatively large. Furthermore, we collect a real-world CML dataset to promote community development. The dataset contains more than 10,000 paired optical and radar images and is continuously being updated. Experimental results on real-world and benchmark datasets validate the effectiveness of the proposed method.

Jiangmeng Li, Wenwen Qiang, Yanan Zhang et al.

As a successful approach to self-supervised learning, contrastive learning aims to learn invariant information shared among distortions of the input sample. While contrastive learning has yielded continuous advancements in sampling strategy and architecture design, it still remains two persistent defects: the interference of task-irrelevant information and sample inefficiency, which are related to the recurring existence of trivial constant solutions. From the perspective of dimensional analysis, we find out that the dimensional redundancy and dimensional confounder are the intrinsic issues behind the phenomena, and provide experimental evidence to support our viewpoint. We further propose a simple yet effective approach MetaMask, short for the dimensional Mask learned by Meta-learning, to learn representations against dimensional redundancy and confounder. MetaMask adopts the redundancy-reduction technique to tackle the dimensional redundancy issue and innovatively introduces a dimensional mask to reduce the gradient effects of specific dimensions containing the confounder, which is trained by employing a meta-learning paradigm with the objective of improving the performance of masked representations on a typical self-supervised task. We provide solid theoretical analyses to prove MetaMask can obtain tighter risk bounds for downstream classification compared to typical contrastive methods. Empirically, our method achieves state-of-the-art performance on various benchmarks.

Zeen Song, Xingzhe Su, Jingyao Wang et al.

In recent years, self-supervised learning (SSL) has emerged as a promising approach for extracting valuable representations from unlabeled data. One successful SSL method is contrastive learning, which aims to bring positive examples closer while pushing negative examples apart. Many current contrastive learning approaches utilize a parameterized projection head. Through a combination of empirical analysis and theoretical investigation, we provide insights into the internal mechanisms of the projection head and its relationship with the phenomenon of dimensional collapse. Our findings demonstrate that the projection head enhances the quality of representations by performing contrastive loss in a projected subspace. Therefore, we propose an assumption that only a subset of features is necessary when minimizing the contrastive loss of a mini-batch of data. Theoretical analysis further suggests that a sparse projection head can enhance generalization, leading us to introduce SparseHead - a regularization term that effectively constrains the sparsity of the projection head, and can be seamlessly integrated with any self-supervised learning (SSL) approaches. Our experimental results validate the effectiveness of SparseHead, demonstrating its ability to improve the performance of existing contrastive methods.

Jiangmeng Li, Hang Gao, Wenwen Qiang et al.

Multi-view representation learning aims to capture comprehensive information from multiple views of a shared context. Recent works intuitively apply contrastive learning to different views in a pairwise manner, which is still scalable: view-specific noise is not filtered in learning view-shared representations; the fake negative pairs, where the negative terms are actually within the same class as the positive, and the real negative pairs are coequally treated; evenly measuring the similarities between terms might interfere with optimization. Importantly, few works study the theoretical framework of generalized self-supervised multi-view learning, especially for more than two views. To this end, we rethink the existing multi-view learning paradigm from the perspective of information theory and then propose a novel information theoretical framework for generalized multi-view learning. Guided by it, we build a multi-view coding method with a three-tier progressive architecture, namely Information theory-guided hierarchical Progressive Multi-view Coding (IPMC). In the distribution-tier, IPMC aligns the distribution between views to reduce view-specific noise. In the set-tier, IPMC constructs self-adjusted contrasting pools, which are adaptively modified by a view filter. Lastly, in the instance-tier, we adopt a designed unified loss to learn representations and reduce the gradient interference. Theoretically and empirically, we demonstrate the superiority of IPMC over state-of-the-art methods.

Hongwei Dong, Fangzhou Han, Lingyu Si et al.

Recent studies have utilized deep learning (DL) techniques to automatically extract features from synthetic aperture radar (SAR) images, which shows great promise for enhancing the performance of SAR automatic target recognition (ATR). However, our research reveals a previously overlooked issue: SAR images to be recognized include not only the foreground (i.e., the target), but also a certain size of the background area. When a DL-model is trained exclusively on foreground data, its recognition performance is significantly superior to a model trained on original data that includes both foreground and background. This suggests that the presence of background impedes the ability of the DL-model to learn additional semantic information about the target. To address this issue, we construct a structural causal model (SCM) that incorporates the background as a confounder. Based on the constructed SCM, we propose a causal intervention based regularization method to eliminate the negative impact of background on feature semantic learning and achieve background debiased SAR-ATR. The proposed causal interventional regularizer can be integrated into any existing DL-based SAR-ATR models to mitigate the impact of background interference on the feature extraction and recognition accuracy. Experimental results on the Moving and Stationary Target Acquisition and Recognition (MSTAR) dataset indicate that the proposed method can enhance the efficiency of existing DL-based methods in a plug-and-play manner.

Jingyao Wang, Wenwen Qiang, Changwen Zheng et al.

Meta-learning seeks to learn a well-generalized model initialization from training tasks to solve unseen tasks. From the "learning to learn" perspective, the quality of the initialization is modeled with one-step gradient decent in the inner loop. However, contrary to theoretical expectations, our empirical analysis reveals that this may expose meta-learning to underfitting. To bridge the gap between theoretical understanding and practical implementation, we reconsider meta-learning from the "Learning" lens. We propose that the meta-learning model comprises two interrelated components: parameters for model initialization and a meta-layer for task-specific fine-tuning. These components will lead to the risks of overfitting and underfitting depending on tasks, and their solutions, fewer parameters vs. more meta-layer, are often in conflict. To address this, we aim to regulate the task information the model receives without modifying the data or model structure. Our theoretical analysis indicates that models adapted to different tasks can mutually reinforce each other, highlighting the effective information. Based on this insight, we propose TRLearner, a plug-and-play method that leverages task relation to calibrate meta-learning. It first extracts task relation matrices and then applies relation-aware consistency regularization to guide optimization. Extensive theoretical and empirical evaluations demonstrate its effectiveness.

Jingyao Wang, Siyu Zhao, Wenwen Qiang et al.

Multi-Modal Learning (MML) aims to learn effective representations across modalities for accurate predictions. Existing methods typically focus on modality consistency and specificity to learn effective representations. However, from a causal perspective, they may lead to representations that contain insufficient and unnecessary information. To address this, we propose that effective MML representations should be causally sufficient and necessary. Considering practical issues like spurious correlations and modality conflicts, we relax the exogeneity and monotonicity assumptions prevalent in prior works and explore the concepts specific to MML, i.e., Causal Complete Cause $C^3$. We begin by defining $C^3$, which quantifies the probability of representations being causally sufficient and necessary. We then discuss the identifiability of $C^3$ and introduce an instrumental variable to support identifying $C^3$ with non-exogeneity and non-monotonicity. Building on this, we conduct the $C^3$ measurement, i.e., \(C^3\) risk. We propose a twin network to estimate it through (i) the real-world branch: utilizing the instrumental variable for sufficiency, and (ii) the hypothetical-world branch: applying gradient-based counterfactual modeling for necessity. Theoretical analyses confirm its reliability. Based on these results, we propose $C^3$ Regularization, a plug-and-play method that enforces the causal completeness of the learned representations by minimizing $C^3$ risk. Extensive experiments demonstrate its effectiveness.

Peizheng Guo, Jingyao Wang, Changwen Zheng et al.

Vision-Language-Action (VLA) models show promising ability in language-guided robotic tasks. However, making VLA policies reliable remains challenging, because a manipulation task is completed through closed-loop interaction, where each action affects subsequent execution. To analyze this problem, we revisit VLA policy during execution and argue that a VLA policy acts both as a planner, which makes task-oriented decisions that change the direction of execution, and as an executor, which realizes these decisions through dense continuous actions. This view suggests that improving VLA reliability requires particular attention to planning actions. Existing optimization methods can imitate actions or improve complete trajectories, but they usually do not explicitly identify planning actions or measure their importance for task success. To address this issue, we propose Planning-Aware Policy Optimization for VLA models (PAPO-VLA). PAPO-VLA first identifies planning actions by jointly considering action variation and trajectory outcome, then estimates their importance through causal sufficiency and causal necessity, and finally incorporates this importance into GRPO advantage estimation. In this way, more important planning actions receive stronger optimization emphasis, while the whole trajectory is still optimized by trajectory-level feedback. Experiments on multiple benchmarks demonstrate the effectiveness of PAPO-VLA.

Zeen Song, Wenwen Qiang, Changwen Zheng et al.

Self-supervised learning (SSL) has recently shown notable success in various visual tasks. However, in terms of discriminability, SSL is still not on par with supervised learning (SL). This paper identifies a key issue, the ``crowding problem," where features from different classes are not well-separated, and there is high intra-class variance. In contrast, SL ensures clear class separation. Our analysis reveals that SSL objectives do not adequately constrain the relationships between samples and their augmentations, leading to poorer performance in complex tasks. We further establish a theoretical framework that connects SSL objectives to cross-entropy risk bounds, explaining how reducing intra-class variance and increasing inter-class separation can improve generalization. To address this, we propose the Dynamic Semantic Adjuster (DSA), a learnable regulator that enhances feature aggregation and separation while being robust to outliers. Comprehensive experiments conducted on diverse benchmark datasets validate that DSA leads to substantial gains in SSL performance, narrowing the performance gap with SL.

Jingyao Wang, Zeen Song, Wenwen Qiang et al.

The long-term goal of machine learning is to learn general visual representations from a small amount of data without supervision, mimicking three advantages of human cognition: i) no need for labels, ii) robustness to data scarcity, and iii) learning from experience. Self-supervised learning and meta-learning are two promising techniques to achieve this goal, but they both only partially capture the advantages and fail to address all the problems. Self-supervised learning struggles to overcome the drawbacks of data scarcity, while ignoring prior knowledge that can facilitate learning and generalization. Meta-learning relies on supervised information and suffers from a bottleneck of insufficient learning. To address these issues, we propose a novel Bootstrapped Meta Self-Supervised Learning (BMSSL) framework that aims to simulate the human learning process. We first analyze the close relationship between meta-learning and self-supervised learning. Based on this insight, we reconstruct tasks to leverage the strengths of both paradigms, achieving advantages i and ii. Moreover, we employ a bi-level optimization framework that alternates between solving specific tasks with a learned ability (first level) and improving this ability (second level), attaining advantage iii. To fully harness its power, we introduce a bootstrapped target based on meta-gradient to make the model its own teacher. We validate the effectiveness of our approach with comprehensive theoretical and empirical study.

Huijie Guo, Jingyao Wang, Lingyu Si et al.

Recent advances in Multimodal Large Language Models (MLLMs) have shown impressive reasoning capabilities across vision-language tasks, yet still face the challenge of compute-difficulty mismatch. Through empirical analyses, we identify that existing decoding methods may waste compute on easy cases while underserving hard ones, affecting both model effectiveness and efficiency. To address this issue, we first develop a theoretical framework that links sampling coverage, instance difficulty, and residual risk. Our analysis reveals that multimodal reasoning exhibits a heavy-tailed difficulty distribution; a small subset of hard or ambiguous samples dominates the residual failure probability. Based on this insight, we propose Coverage-Aware Multimodal Decoding (CAMD), an adaptive inference mechanism that dynamically allocates computation according to estimated uncertainty. CAMD integrates evidence-weighted scoring, posterior coverage estimation, and sequential Bayesian updating to balance efficiency and reliability under a limited token budget. Experiments on various benchmark datasets and baselines demonstrate the effectiveness and advantages of our approach.

Zeen Song, Zihao Ma, Wenwen Qiang et al.

Inference-time reasoning scaling has significantly advanced the capabilities of Large Language Models (LLMs) in complex problem-solving. A prevalent approach involves external search guided by Process Reward Models (PRMs). However, a fundamental limitation of this framework is the epistemic uncertainty of PRMs when evaluating reasoning paths that deviate from their training distribution. In this work, we conduct a systematic analysis of this challenge. We first provide empirical evidence that PRMs exhibit high uncertainty and unreliable scoring on out-of-distribution (OOD) samples. We then establish a theoretical framework proving that while standard search incurs linear regret accumulation, an uncertainty-aware strategy can achieve sublinear regret. Motivated by these findings, we propose Uncertainty-Aware Tree Search (UATS), a unified method that estimates uncertainty via Monte Carlo Dropout and dynamically allocates compute budget using a reinforcement learning-based controller. Extensive experiments demonstrate that our approach effectively mitigates the impact of OOD errors.

Yao Zhou, Zeen Song, Wenwen Qiang et al.

Safety alignment mechanisms in Large Language Models (LLMs) often operate as latent internal states, obscuring the model's inherent capabilities. Building on this observation, we model the safety mechanism as an unobserved confounder from a causal perspective. Then, we propose the Causal Front-Door Adjustment Attack (CFA{$^2$}) to jailbreak LLM, which is a framework that leverages Pearl's Front-Door Criterion to sever the confounding associations for robust jailbreaking. Specifically, we employ Sparse Autoencoders (SAEs) to physically strip defense-related features, isolating the core task intent. We further reduce computationally expensive marginalization to a deterministic intervention with low inference complexity. Experiments demonstrate that CFA{$^2$} achieves state-of-the-art attack success rates while offering a mechanistic interpretation of the jailbreaking process.

Jingyao Wang, Peizheng Guo, Wenwen Qiang et al.

Large language models (LLMs) excel at complex tasks with advances in reasoning capabilities. However, existing reward mechanisms remain tightly coupled to final correctness and pay little attention to the underlying reasoning process: trajectories with sound reasoning but wrong answers receive low credit, while lucky guesses with flawed logic may be highly rewarded, affecting reasoning generalization. From a causal perspective, we interpret multi-candidate reasoning for a fixed question as a family of counterfactual experiments with theoretical supports. Building on this, we propose Group Causal Counterfactual Policy Optimization to explicitly train LLMs to learn generalizable reasoning patterns. It proposes an episodic causal counterfactual reward that jointly captures (i) robustness, encouraging the answer distribution induced by a reasoning step to remain stable under counterfactual perturbations; and (ii) effectiveness, enforcing sufficient variability so that the learned reasoning strategy can transfer across questions. We then construct token-level advantages from this reward and optimize the policy, encouraging LLMs to favor reasoning patterns that are process-valid and counterfactually robust. Extensive experiments on diverse benchmarks demonstrate its advantages.

Wenwen Qiang, Ziyin Gu, Jiahuan Zhou et al.

Training stability remains a critical bottleneck for Group Relative Policy Optimization (GRPO), often manifesting as a trade-off between reasoning plasticity and general capability retention. We identify a root cause as the geometric conflict between plasticity and stability gradients, which leads to destructive interference. Crucially, we argue that deterministic projection methods are suboptimal for GRPO as they overlook the intrinsic stochasticity of group-based gradient estimates. To address this, we propose Probabilistic Conflict Resolution (PCR), a Bayesian framework that models gradients as random variables. PCR dynamically arbitrates conflicts via an uncertainty-aware ``soft projection'' mechanism, optimizing the signal-to-noise ratio. Extensive experiments demonstrate that PCR significantly smooths the training trajectory and achieves superior performance in various reasoning tasks.

Shuyi Zhou, Zeen Song, Wenwen Qiang et al.

Large Language Models remain vulnerable to adversarial prefix attacks (e.g., ``Sure, here is'') despite robust standard safety. We diagnose this vulnerability as Shallow Safety Alignment, stemming from a pathology we term semantic representation decay: as the model generates compliant prefixes, its internal malicious intent signal fades. To address this, we propose Two-Stage Causal-GRPO (TSC-GRPO), a framework designed to achieve intent pinning. First, grounded in causal identifiability theory, we train a causal intent probe to disentangle invariant intent from stylistic perturbations. Second, we internalize this causal awareness into the policy via Group Relative Policy Optimization. By employing a cumulative causal penalty within ``fork-in-the-road'' training scenarios, we force the model to learn that accumulating harmful tokens monotonically decreases reward, enabling robust late-stage refusals. Experiments show that TSC-GRPO significantly outperforms baselines in defending against jailbreak attacks while preserving general utility.

Jiangmeng Li, Yifan Jin, Hang Gao et al.

Graph contrastive learning (GCL) aims to align the positive features while differentiating the negative features in the latent space by minimizing a pair-wise contrastive loss. As the embodiment of an outstanding discriminative unsupervised graph representation learning approach, GCL achieves impressive successes in various graph benchmarks. However, such an approach falls short of recognizing the topology isomorphism of graphs, resulting in that graphs with relatively homogeneous node features cannot be sufficiently discriminated. By revisiting classic graph topology recognition works, we disclose that the corresponding expertise intuitively complements GCL methods. To this end, we propose a novel hierarchical topology isomorphism expertise embedded graph contrastive learning, which introduces knowledge distillations to empower GCL models to learn the hierarchical topology isomorphism expertise, including the graph-tier and subgraph-tier. On top of this, the proposed method holds the feature of plug-and-play, and we empirically demonstrate that the proposed method is universal to multiple state-of-the-art GCL models. The solid theoretical analyses are further provided to prove that compared with conventional GCL methods, our method acquires the tighter upper bound of Bayes classification error. We conduct extensive experiments on real-world benchmarks to exhibit the performance superiority of our method over candidate GCL methods, e.g., for the real-world graph representation learning experiments, the proposed method beats the state-of-the-art method by 0.23% on unsupervised representation learning setting, 0.43% on transfer learning setting. Our code is available at https://github.com/jyf123/HTML.

Zeen Song, Wenwen Qiang, Changwen Zheng et al.

Video contrastive learning (V-CL) has emerged as a popular framework for unsupervised video representation learning, demonstrating strong results in tasks such as action classification and detection. Yet, to harness these benefits, it is critical for the learned representations to fully capture both static and dynamic semantics. However, our experiments show that existing V-CL methods fail to effectively learn either type of feature. Through a rigorous theoretical analysis based on the Structural Causal Model and gradient update, we find that in a given dataset, certain static semantics consistently co-occur with specific dynamic semantics. This phenomenon creates spurious correlations between static and dynamic semantics in the dataset. However, existing V-CL methods do not differentiate static and dynamic similarities when computing sample similarity. As a result, learning only one type of semantics is sufficient for the model to minimize the contrastive loss. Ultimately, this causes the V-CL pre-training process to prioritize learning the easier-to-learn semantics. To address this limitation, we propose Bi-level Optimization with Decoupling for Video Contrastive Learning. (BOD-VCL). In BOD-VCL, we model videos as linear dynamical systems based on Koopman theory. In this system, all frame-to-frame transitions are represented by a linear Koopman operator. By performing eigen-decomposition on this operator, we can separate time-variant and time-invariant components of semantics, which allows us to explicitly separate the static and dynamic semantics in the video. By modeling static and dynamic similarity separately, both types of semantics can be fully exploited during the V-CL training process. BOD-VCL can be seamlessly integrated into existing V-CL frameworks, and experimental results highlight the significant improvements achieved by our method.

Xingzhe Su, Daixi Jia, Fengge Wu et al.

Diffusion Models are a potent class of generative models capable of producing high-quality images. However, they often inadvertently favor certain data attributes, undermining the diversity of generated images. This issue is starkly apparent in skewed datasets like CelebA, where the initial dataset disproportionately favors females over males by 57.9%, this bias amplified in generated data where female representation outstrips males by 148%. In response, we propose a plug-and-play method named Manifold Guidance Sampling, which is also the first unsupervised method to mitigate bias issue in DDPMs. Leveraging the inherent structure of the data manifold, this method steers the sampling process towards a more uniform distribution, effectively dispersing the clustering of biased data. Without the need for modifying the existing model or additional training, it significantly mitigates data bias and enhances the quality and unbiasedness of the generated images.

Jingyao Wang, Yuxuan Yang, Wenwen Qiang et al.

Meta-learning, also known as ``learning to learn'', enables models to acquire great generalization abilities by learning from various tasks. Recent advancements have made these models applicable across various fields without data constraints, offering new opportunities for general artificial intelligence. However, applying these models can be challenging due to their often task-specific, standalone nature and the technical barriers involved. To address this challenge, we develop AwesomeMeta+, a prototyping and learning system designed to standardize the key components of meta-learning within the context of systems engineering. It standardizes different components of meta-learning and uses a building block metaphor to assist in model construction. By employing a modular, building-block approach, AwesomeMeta+ facilitates the construction of meta-learning models that can be adapted and optimized for specific application needs in real-world systems. The system is developed to support the full lifecycle of meta-learning system engineering, from design to deployment, by enabling users to assemble compatible algorithmic modules. We evaluate AwesomeMeta+ through feedback from 50 researchers and a series of machine-based tests and user studies. The results demonstrate that AwesomeMeta+ enhances users' understanding of meta-learning principles, accelerates system engineering processes, and provides valuable decision-making support for efficient deployment of meta-learning systems in complex application scenarios.

Jianqi Zhang, Wenwen Qiang, Jingyao Wang et al.

Transformer-based methods have achieved state-of-the-art performance in time series forecasting (TSF) by capturing positional and semantic topological relationships among input tokens. However, it remains unclear whether existing Transformers fully leverage the intrinsic topological structure among tokens throughout intermediate layers. Through empirical and theoretical analyses, we identify that current Transformer architectures progressively degrade the original positional and semantic topology of input tokens as the network deepens, thus limiting forecasting accuracy. Furthermore, our theoretical results demonstrate that explicitly enforcing preservation of these topological structures within intermediate layers can tighten generalization bounds, leading to improved forecasting performance. Motivated by these insights, we propose the Topology Enhancement Method (TEM), a novel Transformer-based TSF method that explicitly and adaptively preserves token-level topology. TEM consists of two core modules: 1) the Positional Topology Enhancement Module (PTEM), which injects learnable positional constraints to explicitly retain original positional topology; 2) the Semantic Topology Enhancement Module (STEM), which incorporates a learnable similarity matrix to preserve original semantic topology. To determine optimal injection weights adaptively, TEM employs a bi-level optimization strategy. The proposed TEM is a plug-and-play method that can be integrated with existing Transformer-based TSF methods. Extensive experiments demonstrate that integrating TEM with a variety of existing methods significantly improves their predictive performance, validating the effectiveness of explicitly preserving original token-level topology. Our code is publicly available at: \href{https://github.com/jlu-phyComputer/TEM}{https://github.com/jlu-phyComputer/TEM}.

Jingyao Wang, Huijie Guo, Wenwen Qiang et al.

Humans excel at adapting perceptions and actions to diverse environments, enabling efficient interaction with the external world. This adaptive capability relies on the biological nervous system (BNS), which activates different brain regions for distinct tasks. Meta-learning similarly trains machines to handle multiple tasks but relies on a fixed network structure, not as flexible as BNS. To investigate the role of flexible network structure (FNS) in meta-learning, we conduct extensive empirical and theoretical analyses, finding that model performance is tied to structure, with no universally optimal pattern across tasks. This reveals the crucial role of FNS in meta-learning, ensuring meta-learning to generate the optimal structure for each task, thereby maximizing the performance and learning efficiency of meta-learning. Motivated by this insight, we propose to define, measure, and model FNS in meta-learning. First, we define that an effective FNS should possess frugality, plasticity, and sensitivity. Then, to quantify FNS in practice, we present three measurements for these properties, collectively forming the \emph{structure constraint} with theoretical supports. Building on this, we finally propose Neuromodulated Meta-Learning (NeuronML) to model FNS in meta-learning. It utilizes bi-level optimization to update both weights and structure with the structure constraint. Extensive theoretical and empirical evaluations demonstrate the effectiveness of NeuronML on various tasks. Code is publicly available at \href{https://github.com/WangJingyao07/NeuronML}{https://github.com/WangJingyao07/NeuronML}.

Jingyao Wang, Yunhan Tian, Yuxuan Yang et al.

Micro-expressions (MEs) are involuntary movements revealing people's hidden feelings, which has attracted numerous interests for its objectivity in emotion detection. However, despite its wide applications in various scenarios, micro-expression recognition (MER) remains a challenging problem in real life due to three reasons, including (i) data-level: lack of data and imbalanced classes, (ii) feature-level: subtle, rapid changing, and complex features of MEs, and (iii) decision-making-level: impact of individual differences. To address these issues, we propose a dual-branch meta-auxiliary learning method, called LightmanNet, for fast and robust micro-expression recognition. Specifically, LightmanNet learns general MER knowledge from limited data through a dual-branch bi-level optimization process: (i) In the first level, it obtains task-specific MER knowledge by learning in two branches, where the first branch is for learning MER features via primary MER tasks, while the other branch is for guiding the model obtain discriminative features via auxiliary tasks, i.e., image alignment between micro-expressions and macro-expressions since their resemblance in both spatial and temporal behavioral patterns. The two branches of learning jointly constrain the model of learning meaningful task-specific MER knowledge while avoiding learning noise or superficial connections between MEs and emotions that may damage its generalization ability. (ii) In the second level, LightmanNet further refines the learned task-specific knowledge, improving model generalization and efficiency. Extensive experiments on various benchmark datasets demonstrate the superior robustness and efficiency of LightmanNet.

Huijie Guo, Ying Ba, Jie Hu et al.

Self-Supervised Learning (SSL) methods harness the concept of semantic invariance by utilizing data augmentation strategies to produce similar representations for different deformations of the same input. Essentially, the model captures the shared information among multiple augmented views of samples, while disregarding the non-shared information that may be beneficial for downstream tasks. To address this issue, we introduce a module called CompMod with Meta Comprehensive Regularization (MCR), embedded into existing self-supervised frameworks, to make the learned representations more comprehensive. Specifically, we update our proposed model through a bi-level optimization mechanism, enabling it to capture comprehensive features. Additionally, guided by the constrained extraction of features using maximum entropy coding, the self-supervised learning model learns more comprehensive features on top of learning consistent features. In addition, we provide theoretical support for our proposed method from information theory and causal counterfactual perspective. Experimental results show that our method achieves significant improvement in classification, object detection and instance segmentation tasks on multiple benchmark datasets.

Wenwen Qiang, Jingyao Wang, Zeen Song et al.

In this paper, we focus on the out-of-distribution (OOD) generalization of self-supervised learning (SSL). By analyzing the mini-batch construction during the SSL training phase, we first give one plausible explanation for SSL having OOD generalization. Then, from the perspective of data generation and causal inference, we analyze and conclude that SSL learns spurious correlations during the training process, which leads to a reduction in OOD generalization. To address this issue, we propose a post-intervention distribution (PID) grounded in the Structural Causal Model. PID offers a scenario where the spurious variable and label variable is mutually independent. Besides, we demonstrate that if each mini-batch during SSL training satisfies PID, the resulting SSL model can achieve optimal worst-case OOD performance. This motivates us to develop a batch sampling strategy that enforces PID constraints through the learning of a latent variable model. Through theoretical analysis, we demonstrate the identifiability of the latent variable model and validate the effectiveness of the proposed sampling strategy. Experiments conducted on various downstream OOD tasks demonstrate the effectiveness of the proposed sampling strategy.

Xingyu Zhang, Wenwen Qiang, Siyu Zhao et al.

Most existing multivariate time series forecasting methods adopt an all-to-all paradigm that feeds all variable histories into a unified model to predict their future values without distinguishing their individual roles. However, this undifferentiated paradigm makes it difficult to identify variable-specific causal influences and often entangles causally relevant information with spurious correlations. To address this limitation, we propose an all-to-one forecasting paradigm that predicts each target variable separately. Specifically, we first construct a Structural Causal Model from observational data and then, for each target variable, we partition the historical sequence into four sub-segments according to the inferred causal structure: endogenous, direct causal, collider causal, and spurious correlation. The prediction relies solely on the first three causally relevant sub-segments, while the spurious correlation sub-segment is excluded. Furthermore, we propose Causal Informed Transformer (CAIFormer), a novel forecasting model comprising three components: Endogenous Sub-segment Prediction Block, Direct Causal Sub-segment Prediction Block, and Collider Causal Sub-segment Prediction Block, which process the endogenous, direct causal, and collider causal sub-segments, respectively. Their outputs are then combined to produce the final prediction. Extensive experiments on multiple benchmark datasets demonstrate the effectiveness of the CAIFormer.

Jingyao Wang, Wenwen Qiang, Zeen Song et al.

Large language models (LLMs) excel at complex tasks thanks to advances in their reasoning abilities. However, existing methods overlook the trade-off between reasoning effectiveness and efficiency, often encouraging unnecessarily long reasoning chains and wasting tokens. To address this, we propose Learning to Think (L2T), an information-theoretic reinforcement fine-tuning framework for LLMs to make the models achieve optimal reasoning with fewer tokens. Specifically, L2T treats each query-response interaction as a hierarchical session of multiple episodes and proposes a universal dense process reward, i.e., quantifies the episode-wise information gain in parameters, requiring no extra annotations or task-specific evaluators. We propose a method to quickly estimate this reward based on PAC-Bayes bounds and the Fisher information matrix. Theoretical analyses show that it significantly reduces computational complexity with high estimation accuracy. By immediately rewarding each episode's contribution and penalizing excessive updates, L2T optimizes the model via reinforcement learning to maximize the use of each episode and achieve effective updates. Empirical results on various reasoning benchmarks and base models demonstrate the advantage of L2T across different tasks, boosting both reasoning effectiveness and efficiency.

Zeen Song, Siyu Zhao, Xingyu Zhang et al.

Contrastive Language-Image Pretraining (CLIP) has achieved remarkable success, but its performance can degrade when fine-tuned in out-of-distribution (OOD) scenarios. We model the prediction process using a Structural Causal Model (SCM) and show that the causal mechanism involving both invariant and variant factors in training environments differs from that in test environments. In contrast, the causal mechanism with solely invariant factors remains consistent across environments. We theoretically prove the existence of a linear mapping from CLIP embeddings to invariant factors, which can be estimated using interventional data. Additionally, we provide a condition to guarantee low OOD risk of the invariant predictor. Based on these insights, we propose the Invariant Causal Mechanism of CLIP (CLIP-ICM) framework. CLIP-ICM involves collecting interventional data, estimating a linear projection matrix, and making predictions within the invariant subspace. Experiments on several OOD datasets show that CLIP-ICM significantly improves the performance of CLIP. Our method offers a simple but powerful enhancement, boosting the reliability of CLIP in real-world applications.

Wenwen Qiang, Jingyao Wang, Changwen Zheng et al.

In this paper, we investigate what constitutes a good representation or model in self-supervised learning (SSL). We argue that a good representation should exhibit universality, characterized by three essential properties: discriminability, generalizability, and transferability. While these capabilities are implicitly desired in most SSL frameworks, existing methods lack an explicit modeling of universality, and its theoretical foundations remain underexplored. To address these gaps, we propose General SSL (GeSSL), a novel framework that explicitly models universality from three complementary dimensions: the optimization objective, the parameter update mechanism, and the learning paradigm. GeSSL integrates a bi-level optimization structure that jointly models task-specific adaptation and cross-task consistency, thereby capturing all three aspects of universality within a unified SSL objective. Furthermore, we derive a theoretical generalization bound, ensuring that the optimization process of GeSSL consistently leads to representations that generalize well to unseen tasks. Empirical results on multiple benchmark datasets demonstrate that GeSSL consistently achieves superior performance across diverse downstream tasks, validating its effectiveness in modeling universal representations.

Jiangmeng Li, Zehua Zang, Qirui Ji et al.

Representations learned by self-supervised approaches are generally considered to possess sufficient generalizability and discriminability. However, we disclose a nontrivial mutual-exclusion relationship between these critical representation properties through an exploratory demonstration on self-supervised learning. State-of-the-art self-supervised methods tend to enhance either generalizability or discriminability but not both simultaneously. Thus, learning representations jointly possessing strong generalizability and discriminability presents a specific challenge for self-supervised learning. To this end, we revisit the learning paradigm of self-supervised learning from the perspective of evolutionary game theory (EGT) and outline the theoretical roadmap to achieve a desired trade-off between these representation properties. EGT performs well in analyzing the trade-off point in a two-player game by utilizing dynamic system modeling. However, the EGT analysis requires sufficient annotated data, which contradicts the principle of self-supervised learning, i.e., the EGT analysis cannot be conducted without the annotations of the specific target domain for self-supervised learning. Thus, to enhance the methodological generalization, we propose a novel self-supervised learning method that leverages advancements in reinforcement learning to jointly benefit from the general guidance of EGT and sequentially optimize the model to chase the consistent improvement of generalizability and discriminability for specific target domains during pre-training. Theoretically, we establish that the proposed method tightens the generalization error upper bound of self-supervised learning. Empirically, our method achieves state-of-the-art performance on various benchmarks.

Xingyu Zhang, Hanyun Du, Zeen Song et al.

Large Language Models (LLMs) have recently shown exceptional potential in time series forecasting, leveraging their inherent sequential reasoning capabilities to model complex temporal dynamics. However, existing approaches typically employ a naive autoregressive generation strategy. We identify a critical theoretical flaw in this paradigm: during inference, the model operates in an open-loop manner, consuming its own generated outputs recursively. This leads to inevitable error accumulation (exposure bias), where minor early deviations cascade into significant trajectory drift over long horizons. In this paper, we reformulate autoregressive forecasting through the lens of control theory, proposing \textbf{F-LLM} (Feedback-driven LLM), a novel closed-loop framework. Unlike standard methods that passively propagate errors, F-LLM actively stabilizes the trajectory via a learnable residual estimator (Observer) and a feedback controller. Furthermore, we provide a theoretical guarantee that our closed-loop mechanism ensures uniformly bounded error, provided the base model satisfies a local Lipschitz constraint. Extensive experiments demonstrate that F-LLM significantly mitigates error propagation, achieving good performance on time series benchmarks.

Jianqi Zhang, Jingyao Wang, Wenwen Qiang et al.

The World Wide Web needs reliable predictive capabilities to respond to changes in user behavior and usage patterns. Time series forecasting (TSF) is a key means to achieve this goal. In recent years, the large language models (LLMs) for TSF (LLM4TSF) have achieved good performance. However, there is a significant difference between pretraining corpora and time series data, making it hard to guarantee forecasting quality when directly applying LLMs to TSF; fine-tuning LLMs can mitigate this issue, but often incurs substantial computational overhead. Thus, LLM4TSF faces a dual challenge of prediction performance and compute overhead. To address this, we aim to explore a method for improving the forecasting performance of LLM4TSF while freezing all LLM parameters to reduce computational overhead. Inspired by in-context learning (ICL), we propose LVICL. LVICL uses our vector-injected ICL to inject example information into a frozen LLM, eliciting its in-context learning ability and thereby enhancing its performance on the example-related task (i.e., TSF). Specifically, we first use the LLM together with a learnable context vector adapter to extract a context vector from multiple examples adaptively. This vector contains compressed, example-related information. Subsequently, during the forward pass, we inject this vector into every layer of the LLM to improve forecasting performance. Compared with conventional ICL that adds examples into the prompt, our vector-injected ICL does not increase prompt length; moreover, adaptively deriving a context vector from examples suppresses components harmful to forecasting, thereby improving model performance. Extensive experiments demonstrate the effectiveness of our approach.

Huijie Guo, Jingyao Wang, Peizheng Guo et al.

In this paper, we explore the transferability of SSL by addressing two central questions: (i) what is the representation transferability of SSL, and (ii) how can we effectively model this transferability? Transferability is defined as the ability of a representation learned from one task to support the objective of another. Inspired by the meta-learning paradigm, we construct multiple SSL tasks within each training batch to support explicitly modeling transferability. Based on empirical evidence and causal analysis, we find that although introducing task-level information improves transferability, it is still hindered by task conflict. To address this issue, we propose a Task Conflict Calibration (TC$^2$) method to alleviate the impact of task conflict. Specifically, it first splits batches to create multiple SSL tasks, infusing task-level information. Next, it uses a factor extraction network to produce causal generative factors for all tasks and a weight extraction network to assign dedicated weights to each sample, employing data reconstruction, orthogonality, and sparsity to ensure effectiveness. Finally, TC$^2$ calibrates sample representations during SSL training and integrates into the pipeline via a two-stage bi-level optimization framework to boost the transferability of learned representations. Experimental results on multiple downstream tasks demonstrate that our method consistently improves the transferability of SSL models.

Lingyu Si, Jingyao Wang, Wenwen Qiang

Self-supervised contrastive learning (SSCL) has recently demonstrated superiority in multiple downstream tasks. In this paper, we generalize the standard SSCL methods to a Generalized Learning Framework (GLF) consisting of two parts: the aligning part and the constraining part. We analyze three existing SSCL methods: BYOL, Barlow Twins, and SwAV, and show that they can be unified under GLF with different choices of the constraining part. We further propose empirical and theoretical analyses providing two insights into designing the constraining part of GLF: intra-class compactness and inter-class separability, which measure how well the feature space preserves the class information of the inputs. However, since SSCL can not use labels, it is challenging to design a constraining part that satisfies these properties. To address this issue, we consider inducing intra-class compactness and inter-class separability by iteratively capturing the dynamic relationship between anchor and other samples and propose a plug-and-play method called Adaptive Distribution Calibration (ADC) to ensure that samples that are near or far from the anchor point in the original input space are closer or further away from the anchor point in the feature space. Both the theoretical analysis and the empirical evaluation demonstrate the superiority of ADC.

Ziyin Gu, Jingyao Wang, Ran Zuo et al.

Recent advances in large language models (LLMs) have broadened their applicability across diverse tasks, yet specialized domains still require targeted post training. Among existing methods, Group Relative Policy Optimization (GRPO) stands out for its efficiency, leveraging groupwise relative rewards while avoiding costly value function learning. However, GRPO treats candidate responses as independent, overlooking semantic interactions such as complementarity and contradiction. To address this challenge, we first introduce a Structural Causal Model (SCM) that reveals hidden dependencies among candidate responses induced by conditioning on a final integrated output forming a collider structure. Then, our causal analysis leads to two insights: (1) projecting responses onto a causally informed subspace improves prediction quality, and (2) this projection yields a better baseline than query only conditioning. Building on these insights, we propose Group Causal Policy Optimization (GCPO), which integrates causal structure into optimization through two key components: a causally informed reward adjustment and a novel KL regularization term that aligns the policy with a causally projected reference distribution. Comprehensive experimental evaluations demonstrate that GCPO consistently surpasses existing methods, including GRPO across multiple reasoning benchmarks.

Ruike Song, Zeen Song, Huijie Guo et al.

External reasoning systems combine language models with process reward models (PRMs) to select high-quality reasoning paths for complex tasks such as mathematical problem solving. However, these systems are prone to reward hacking, where high-scoring but logically incorrect paths are assigned high scores by the PRMs, leading to incorrect answers. From a causal inference perspective, we attribute this phenomenon primarily to the presence of confounding semantic features. To address it, we propose Causal Reward Adjustment (CRA), a method that mitigates reward hacking by estimating the true reward of a reasoning path. CRA trains sparse autoencoders on the PRM's internal activations to recover interpretable features, then corrects confounding by using backdoor adjustment. Experiments on math solving datasets demonstrate that CRA mitigates reward hacking and improves final accuracy, without modifying the policy model or retraining PRM.