IJCAI 2026 Papers

Adversarial training is effective on balanced datasets, but its robustness degrades under long-tailed class distributions, where tail classes suffer high robust error and unstable decision boundaries. We propose \emph{Manifold-Constrained Adversarial Training (MCAT)}, a unified framework that enforces the semantic validity of adversarial examples by penalizing deviations from class-conditional manifolds in feature space, while promoting balanced geometric separation across classes via an ETF-inspired regularization. We provide theoretical results that link geometric separation to lower bounds on adversarially robust margins, and show that manifold-constrained adversarial risk upper-bounds robust risk on high-density semantic regions. Extensive experiments on standard long-tailed benchmarks demonstrate consistent improvements in overall, balanced, and tail-class adversarial robustness. The codes and appendix are available on https://github.com/yneversky/MCAT.

Symbolic regression (SR) aims to recover compact and interpretable mathematical expressions from data. Genetic programming (GP) directly searches over symbolic structures, but its population dynamics can make it difficult to reliably preserve and accumulate useful subexpressions. In contrast, reinforcement learning (RL)-based SR has shown strong empirical performance, suggesting that learned sampling dynamics may capture useful regularities in symbolic search. Motivated by this contrast, we analyze expressions sampled during RL training and identify a recurring pattern, termed progressive subexpression reuse, where useful simple subexpressions emerge early, become increasingly frequent, and support the formation of more complex structures. Based on this observation, we propose Reinforcement Genetic Programming (RGP), a purely GP-based and non-RL framework that explicitly realizes a stage-wise retention--reintroduction loop through a dynamically maintained subexpression pool and pool-guided population initialization. Experiments on standard SR benchmarks show that RGP matches or outperforms strong baselines without RL policy training, suggesting that progressive subexpression reuse is an effective mechanism for SR search. We release our code at https://github.com/wuxiangdong586-max/RGP.

Federated learning (FL) enables collaborative training of deep models while keeping raw data on local devices. A critical factor affecting the global performance is server-side aggregation. The most common strategy weights client updates by the amount of local data, which is often unreliable as the server cannot inspect data. More importantly, it ignores substantial variations in data quality (e.g., label noise) across heterogeneous clients. To address this issue, we propose a data-Quality-aware aggregation framework by introducing an Evolutionary-computation-inspired design into Federated learning (FedEvoQ), with a lightweight dual-branch architecture. Each client trains a main-task branch as usual, while a frozen auxiliary branch (EvoQ) performs evolutionary-style controlled sampling with crossover and mutation at the input and representation level to probe data diversity and reliability. EvoQ outputs a compact quality score that jointly reflects learnability (loss-based signal), geometric consistency (stability under perturbations), representation diversity, and out-of-distribution (OOD) deviation, enabling the server to compute quality-only aggregation weights. Extensive experiments on CIFAR10, CIFAR100, and PathMNIST under noisy and Non-independent and identically distributed (Non-IID) settings demonstrate that FedEvoQ consistently outperforms 10 state-of-the-art (SOTA) federated aggregation methods, improving the average test accuracy of the global model by over 2% with modest overhead.

Partial Dependence Plots (PDPs) visualize how changes in a single feature affect the average model prediction. They are widely used in practice to interpret decision tree ensembles and other machine learning models. Joint-PDPs extend this idea to pairs of features, revealing their combined effect. Partial Dependence Interaction Values (PDIVs) measure feature interactions. The Any-Order-PDIVs task computes these interactions for every feature subset across all rows of the dataset. We introduce WOODELF++, a unified and efficient approach for computing all these useful explainability tools on decision tree ensembles, building on WOODELF, an algorithm for efficient SHAP computation. By deriving suitable metrics over pseudo-Boolean functions, WOODELF++ can compute PDPs (exact and approximate), Joint-PDPs, and Any-Order-PDIVs in a unified framework. Our method delivers substantial complexity improvements over the state of the art, including an exponential gain for Any-Order-PDIVs. Additionally, we introduce and efficiently compute Full PDPs, which leverage the model’s split thresholds to faithfully capture its behavior across all possible feature values. WOODELF++ is implemented in pure Python and supports GPU acceleration. On a dataset with 400,000 rows, WOODELF++ computes PDP and Joint-PDP up to 6x faster than the state of the art and up to five orders of magnitude faster than scikit-learn. For Any-Order-PDIVs, the gap is even larger: WOODELF++ computes all interaction values in 5 minutes, while the state of the art is estimated to require over 1,000,000 years.

This paper addresses the problem of learning to sparsify stochastic linear bandits, where a decision-maker sequentially selects actions from a high-dimensional space subject to a sparsity constraint on the number of nonzero elements in the action vector. The key challenge lies in minimizing cumulative regret while tackling the potential NP-hardness of finding optimal sparse actions due to the inherent combinatorial structure of the problem. We propose an adaptively phased exploration and exploitation algorithmic framework, utilizing ordinary least squares for parameter learning and specialized subroutines for sparse action selection. When the action set is a Euclidean ball, optimal sparse actions can be efficiently computed, enabling us to establish a Õ(d √T) regret, where d is the dimension of the action vector and T is the time horizon length. For general convex and compact action sets where finding optimal sparse actions is intractable, we employ a greedy subroutine. For general strongly convex action sets, we derive a Õ(d √T) α-regret; for general compact sets lacking strong convexity, we establish a Õ(d T^(2/3)) α-regret, where α pertains to the approximation ratio of the greedy algorithm. Finally, we validate the performance of our algorithms using extensive experiments.

Time series forecasting (TSF) refers to a fundamental task of predicting future sequential data based on historical observations. One representative category of TSF methods is transformer-based approaches, which translate time series into token sequences (i.e., as raw texts) before applying well-established transformers. In this paper, we conducted extensive preliminary experiments to evaluate their stability against various noises, and we empirically observed an interesting phenomenon: applying noises with appropriate levels to training time series can promote the predictive accuracy of transformer-based TSF methods. We analyze this phenomenon from the perspective of token attention and find one basic reason is that applying noises with appropriate levels can lead to token attention weights wavy, which is more consistent to the basic characteristic of time series data. Motivated by this phenomenon and analysis, we apply perturbations and propose a sub-objective wrt, perturbations constraining token attention weights to be wavy. Accordingly, we propose a novel TSF method, namely Wave-Attention-aware TransformER (WATER). We empirically evaluate \baby across the commonly used benchmark datasets, and experimental results indicate that it consistently outperforms the existing TSF methods.

Time-series clustering remains challenging due to the inherent trade-off between clustering effectiveness and computational efficiency. Similarity-based methods often suffer from quadratic complexity caused by pairwise distance computations, while deep learning–based approaches typically rely on costly iterative training and a large number of trainable parameters. In this paper, we propose MSRGC-Net, an efficient time-series clustering framework that integrates multiscale reservoir computing, granular-ball-based anchoring graph construction, and consensus learning. MSRGC-Net adopts a training-free reservoir computing paradigm to extract multiscale temporal representations from raw time series without backpropagation, significantly reducing computational overhead. To capture the intrinsic structure of the resulting representations, granular-ball computing is employed to adaptively model data distributions via density-consistent regions, yielding compact and robust anchor graph representations. Furthermore, a consensus-based anchoring graph optimization strategy is introduced to effectively align multiscale reservoir representations and integrate complementary information across temporal scales. Extensive experiments on widely used univariate and multivariate benchmark datasets demonstrate that MSRGC-Net consistently outperforms state-of-the-art methods in clustering performance while maintaining superior computational efficiency.

Unsupervised Domain Adaptation for Semantic Segmentation (UDA-SS) has seen significant progress in recent years. Existing UDA-SS approaches mostly adopt a pseudo-labeling schema to adapt model in the target domain, but they often overlook the inherent long-tailed data distribution in segmentation. We find that such scarce tail samples can lead to representation collapse for tail classes and further hinder the quality of pseudo-labels. To address these challenges, we propose MarCon, a framework that mitigates the long-tailed problem from both feature representation and pseudo-labeling perspectives. Specifically, to explicitly learn a Maximum-Margin Distribution, we derive a reformulated pixel-level contrastive learning objective by modeling feature distributions with the von Mises-Fisher (vMF) distribution. It enforces strict margins to enhance intra-class compactness and inter-class separability, preventing tail classes from being overwhelmed by head classes. Furthermore, to mitigate label noise, we introduce a Reliability-aware Filter (RaF) based on the vMF-derived metrics, which performs adaptive class-wise pixel reliability assessment to identify unreliable pixels and attenuate their contribution to model training, thereby mitigating confirmation bias. Extensive experiments on GTA → Cityscapes and SYNTHIA → Cityscapes demonstrate that MarCon consistently outperforms current leading methods across various transformer-based architectures.

Low-rank gradient projection (LoRP) has recently emerged as a memory-efficient alternative to low-rank adapters (LoRA) for finetuning large language models. Existing LoRP methods, however, implicitly fix the projection unit to a single gradient row, leaving the effect of grouping multiple rows (or subdividing a row) largely unexplored. In this work, we systematically investigate the impact of the projection unit on LoRP methods. Specifically, we extend existing LoRP approaches by introducing an additional degree of freedom, projection granularity, beyond the traditional rank hyperparameter. This enables a framework capable of performing Various-grained Low-Rank Projection of gradients, which we term VLoRP. Using VLoRP, we observe that, under an identical memory budget, fine-grained projections consistently deliver superior performance. Moreover, VLoRP requires no extra computation and minimal code changes, effectively providing a no-cost accuracy boost to LoRP. Finally, we provide convergence analysis on VLoRP with either SGD or an Adam-based memory-efficient optimizer, and extensive experiments are conducted to validate our findings, covering tasks such as Commonsense Reasoning, MMLU, and GSM8K

As a key technique for improving throughput by increasing clock frequency and reducing Cycles per Instruction (CPI), pipeline design increasingly relies on automated methods with the growing scale of modern circuits. However, existing automated pipelining methods often decouple partitioning from CPI optimization, implicitly assuming that partitioning affects only clock frequency. We observe that different partitioning strategies significantly alter the pipeline stage differences between logic gates, thereby impacting the average number of stall cycles and ultimately affecting CPI. Thus, automated pipeline partitioning should jointly reduce CPI and improve clock frequency while ensuring functional correctness, resulting in a multi-objective optimization problem. To address this issue, we propose EvoPartition, an evolutionary partitioning framework for automated pipeline design. Specifically, EvoPartition first inserts virtual nodes to guarantee functional correctness. Then, through iterative evolution for reducing CPI and improving clock frequency, pipeline stage assignments of all nodes are determined. Experimental results on ISCAS85 and EPFL show that EvoPartition achieves a 7.34% CPI optimization by reducing stage differences by 16.23%, and further improves throughput by 7.22% compared to the state-of-the-art.

Federated Reinforcement Learning from Human Feedback (RLHF) enables the collaborative alignment of Large Language Models (LLMs) while preserving privacy, yet it faces critical bottlenecks arising from data heterogeneity. Existing approaches typically rely on rigid client-level clustering, which overlooks intra-client heterogeneity and fails to adapt to the multifaceted needs of individual users. To address this, we propose FedPrism, a novel framework that shifts alignment granularity from the coarse client level to the precise instance level. Similar to an optical prism dispersing mixed light into distinct spectral components, FedPrism decomposes complex, intra-client heterogeneous data streams by dynamically routing individual samples to specialized experts based on semantic features. Crucially, we devise a Posterior-Guided Performance Distillation (PGPD) mechanism that leverages experts' actual training loss as self-supervision to autonomously refine routing policies without explicit labels. Extensive experiments show that FedPrism not only establishes new state-of-the-art results but also effectively mitigates the negative transfer prevalent in non-IID settings, ensuring superior alignment fidelity.

Federated multi-view clustering (FedMVC) has been widely used to discover latent structures in distributed multi-view data, yet most methods still rely on the restrictive assumption of independent and identically distributed (IID) data. In practice, non-IID distributions with partial and imbalanced categories cause clients to learn biased and local representations, leading to model bias and unstable federated training performance. To address these issues, we propose a FedMVC framework called Bridging Inter-View and Client Heterogeneity: Federated Multi-View Clustering under Non-IID Data (H²-FedMVC), which tackles both inter-view and client heterogeneity in mixed-view settings. We propose a leader-guided learning mechanism to address inter-view heterogeneity by enhancing complementary and discriminative feature learning, and an expert adaptive aggregation strategy to handle client heterogeneity by prioritizing well-matched experts in global updates. Theoretical analysis and experiments demonstrate superior performance in heterogeneous mixed-view FedMVC settings, with code provided at https://github.com/JiazhenWang-0522/H2-FedMVC.

Multi-view learning aims to enhance performance by integrating information from multiple sources. While different views offer complementary perspectives, extracting consistent and discriminative representations remains a significant challenge due to discrepancies in representation and presence of noise. Existing methods typically separate the denoising and alignment processes, mapping the denoised heterogeneous views to a shared subspace for alignment. This means that noise may propagate or even amplify during the alignment stage, ultimately leading to suboptimal solutions. To address this issue, we propose center-guided spectral diffusion, which replaces traditional alignment with generative modeling. This method avoids noise propagation in multi-view alignment process and prevents multi-view features be aligned to the noise subspace during fusion process. Specifically, we first performs unconditional diffusion in both the feature space and a low-rank spectral space to learn a stable consensus centre anchor. This anchor is then used to condition a guided generative diffusion process, enabling the model to generate more consistent and realistic sample representations. By combining conditional and unconditional diffusion, the proposed method alleviates the noise amplification problem commonly found in traditional alignment methods. Experimental results show that the proposed method outperforms state-of-the-art methods on several datasets.

Stochastic gradient methods are central to large-scale learning, yet their generalization theory typically relies on independent sampling assumptions. In many practical applications, data are generated by Markov chains and learning is performed in a decentralized manner, which introduces significant analytical challenges. In this work, we investigate the stability and generalization of decentralized stochastic gradient descent (SGD) and stochastic gradient descent ascent (SGDA) under Markov chain sampling. Leveraging a stability-based framework, we characterize how Markovian dependence and decentralized communication jointly influence generalization behavior. Our analysis captures the effects of network topology, Markov chain mixing properties, and primal–dual dynamics. We establish non-asymptotic generalization bounds for both algorithms, extending existing results on Markov stochastic gradient methods to decentralized and minimax settings.

Multi-view clustering aims to utilize information from multiple feature representations to uncover underlying data structures. Most existing methods emphasize learning a consensus representation by enforcing consistency across views. However, those structures that cannot be directly incorporated into the clustering space receive little attention, and are often discarded as noise in practice. In this paper, we argue that such information can be informative signals for the clustering process and should be explicitly modeled rather than suppressed or ignored. Specifically, we propose salient-residual decoupled multi-view learning for clustering, SRDMVC, introducing a novel decomposition-fusion iterative optimization, which separates the feature space into a salient space and a residual subspace effectively and fuses them using a novel attention mechanism. The residual subspace can capture the deep-level structure of view information, making positive contributions to the clustering process, under proper constraints. Without assuming stronger view alignment or complementarity, SRDMVC enhances cluster discrimination, avoids spurious consensus, and alleviates representation degradation. Extensive experiments on benchmark datasets demonstrate the superior performance of the method we propose.

Many practical decision-making problems involve tasks whose success depends on the entire system history, rather than on achieving a state with desired properties. Markovian Reinforcement Learning (RL) can be inadequate for such tasks, while modeling them as non-Markovian reward decision processes (NMRDPs) enables agents to handle temporal dependencies. Existing approaches offer limited formal guarantees on both (near-)optimality and sample efficiency. We address both issues with QR-max, a novel model-based algorithm for NMRDPs with discrete actions that factorizes Markovian transition learning from non-Markovian reward handling via reward machines. To our knowledge, this is the first model-based RL algorithm for discrete-action NMRDPs that leverages this factorization to obtain PAC convergence to ε-optimal policies with polynomial sample complexity. We then extend QR-max to continuous state spaces with Bucket-QR-max, a SimHash-based discretizer that preserves the same factorized structure and achieves fast and stable learning without manual gridding or function approximation. We experimentally compare our method with state-of-the-art model-based RL approaches on environments of increasing complexity, showing substantially improved sample efficiency and greater robustness in finding optimal policies.

The primary challenge in incremental learning for deep neural networks (DNNs) lies in balancing memory stability with learning plasticity as they incrementally acquire new knowledge. Existing incremental learning methods typically require storing portions of old datasets or relying on specialized network architectures to achieve learning without forgetting. Recent neuroscience findings reveal that hippocampal sharp-wave ripples play critical roles in consolidating neocortex memory by internally triggering the reactivation of spontaneous neural circuits without external stimulus. Inspired by this biological mechanism, we propose an incremental learning method termed Sharp-Wave Ripples Learning (SWRL), a novel incremental learning framework composed of two functionally complementary modules: a neocortex-like network for long-term memory and a hippocampus-like model for rapid new-knowledge encoding. SWRL continuously integrates newly acquired knowledge from the hippocampus-like module into the neocortex-like network via a meta-weighted memory integration mechanism guided by internally generated SWR-like signals. Our framework is task-agnostic, architecture-agnostic, and exemplar-free. Experimental results on multiple public benchmarks demonstrate that SWRL achieves superior incremental learning performance under various scenarios across diverse DNN architectures, requiring no access to previous data or modifications to the base architecture. The code and appendix are both released at https://github.com/SYVAE/SWRL.

Label Distribution Learning (LDL) represents each instance with a label distribution, but these distributions are often incomplete in practice due to high annotation costs and annotators’ cognitive burden. Recent Incomplete Label Distribution Learning (InLDL) methods leverage global and local correlations to impute missing values, imputation errors remain systematic and non-uniform across labels. However, in the more challenging scenario of imbalanced label distributions, such non-uniform errors disproportionately harm low-degree labels, skewing learning toward dominant labels. To address the more complex challenge, we propose Label Distribution Imputation and Bias-Corrective Representations (LDIBR) for incomplete and imbalanced label distribution learning. LDIBR performs instance-adaptive imputation conditioned on instance features and the binary observation mask. It learns a prior, a reliability-gated correction, and entry-wise fusion weights to produce a normalized imputed distribution. Additionally, LDIBR decomposes the imputed distributions into a shared low-rank core and an instance-specific residual. The core captures stable global degree patterns, while the residual isolates sample-specific deviations and imputation noise. Comprehensive experiments across multiple benchmark datasets with a 50% missing ratio validate both the effectiveness and robustness of the LDIBR approach. The code is available at https://github.com/wenhuihji/LDCBR.

3D molecular linker design is a critical task in structure-based drug discovery, which requires the precise synthesis of chemical bridges to connect fragments within the constrained environment of a protein binding pocket. Existing methods often suffer from environmental blindness, treating the inter-fragment space as a vacuum and yielding candidates with poor binding affinity or severe steric clashes. To address this, we propose LinkerBridge, an equivariant framework that unifies biochemical semantics with physical constraints through two innovations: a Contextual Interaction-Aware Representation module that internalizes pre-existing biochemical semantics, and a Differentiable Physical Guidance mechanism derived from Van der Waals potentials to steer generation away from collision zones. Extensive evaluations on ZINC, GEOM, and BindingMOAD benchmarks, as well as Hsp90 and JNK3 case studies, demonstrate that LinkerBridge significantly outperforms state-of-the-art methods in real-world drug discovery.