AAAI 2026 Papers

Current methods for editing pre-trained models face significant challenges, primarily high computational costs and limited scalability. Task arithmetic has recently emerged as a promising solution, using simple arithmetic operations—addition and negation—based on task vectors which are the differences between fine-tuned and pre-trained model weights, to efficiently modify model behavior. However, the full potential of task arithmetic remains underexplored, primarily due to limited mechanisms for overcoming optimization stagnation. To address this challenge, we introduce the notion of difference vector, a generalized form of task vectors derived from the historical movements during optimization. Using difference vectors as directed perturbations, we proposed the Difference Vector-based Anisotropic Scaling Iterative algorithm (DV-BASI) to enable a continuous optimization process for task arithmetic methods without relying on any additional modules or components. Notably, by leveraging escapability and directional advantages of difference vectors, the average performance on different tasks of the multi-task model merged by DV-BASI may even outperform models individually fine-tuned. Based on this observation, we extend the application of difference vectors to a feasible fine-tuning method for single-task models. On the practical side, DV-BASI allows expressive searching directions with few learnable parameters and forms a scalable framework. We also integrate DV-BASI with task arithmetic methods and advanced optimization techniques to achieve state-of-the-art performance on both supervised and unsupervised evaluation protocols.

Recent studies on Neural Collapse (NC) reveal that, under class-balanced conditions, the class feature means and the classifier weights spontaneously align into a simplex equiangular tight frame (ETF). In long-tailed regimes, however, severe sample imbalance tends to prevent the emergence of the NC phenomenon, resulting in poor generalization performance.Current efforts predominantly seek to recover the ETF geometry by imposing constraints on features or classifier weights, yet overlook a critical problem: There is a pronounced misalignment between the feature and the classifier weight spaces. In this paper, we theoretically quantify the harm of such misalignment through an optimal error exponent analysis.Built on this insight, we propose three explicit alignment strategies that plug-and-play into existing long-tail methods without architectural change. Extensive experiments on the CIFAR-10-LT, CIFAR-100-LT, and ImageNet-LT datasets consistently boost examined baselines and achieve the state-of-the-art performances.

Machine learning under limited computational resources has gained increasing attention recently. A common yet challenging scenario is managing multiple time-constrained learning tasks with budgeted computational resources, known as Computational Resource Efficient Learning (CoRE-Learning). To this end, a recently proposed framework, Learning with Adaptive Resource Allocation (LARA), offers a preliminary approach. In this paper, we point out the limitations of LARA, including its reliance on interpolation-based extrapolation methods, the need for a fixed exploration phase, and the use of high-frequency re-estimation and reallocation strategies. To address these issues, we propose Look-ahead and immediate Resource Allocation (LaiRA). Our approach incorporates an efficient Dynamic Kalman Filtering (DKF) for look-ahead feasibility check with limited data and a weight-based online estimator for immediate performance evaluation. For resource allocation, LaiRA constructs an Upper Confidence Bound (UCB) to enable adaptive exploration and introduces an adaptive time-slicing method to reduce task switching costs. Empirical studies validate the effectiveness of our approach.

Spiking Neural Networks (SNNs) offer a promising direction for energy-efficient and brain-inspired computing, yet their vulnerability to adversarial perturbations remains poorly understood. In this work, we revisit the adversarial robustness of SNNs through the lens of temporal ensembling, treating the network as a collection of evolving sub-networks across discrete timesteps. This formulation uncovers two critical but underexplored challenges—the fragility of individual temporal sub-networks and the tendency for adversarial vulnerabilities to transfer across time. To overcome these limitations, we propose Robust Temporal self-Ensemble (RTE), a training framework that improves the robustness of each sub-network while reducing the temporal transferability of adversarial perturbations. RTE integrates both objectives into a unified loss and employs a stochastic sampling strategy for efficient optimization. Extensive experiments across multiple benchmarks demonstrate that RTE consistently outperforms existing training methods in robust-accuracy trade-off. Additional analyses reveal that RTE reshapes the internal robustness landscape of SNNs, leading to more resilient and temporally diversified decision boundaries. Our study highlights the importance of temporal structure in adversarial learning and offers a principled foundation for building robust spiking models.

Time series forecasting relies on predicting future values from historical data, yet most state-of-the-art approaches—including transformer and multilayer perceptron-based models—optimize using Mean Squared Error (MSE), which has two fundamental weaknesses: its point-wise error computation fails to capture temporal relationships, and it does not account for inherent noise in the data. To overcome these limitations, we introduce the Residual-Informed Loss (RI-Loss), a novel objective function based on the Hilbert-Schmidt Independence Criterion (HSIC). RI-Loss explicitly models noise structure by enforcing dependence between the residual sequence and a random time series, enabling more robust, noise-aware representations. Theoretically, we derive the first non-asymptotic HSIC bound with explicit double-sample complexity terms, achieving optimal convergence rates through Bernstein-type concentration inequalities and Rademacher complexity analysis. This provides rigorous guarantees for RI-Loss optimization while precisely quantifying kernel space interactions. Empirically, experiments across eight real-world benchmarks and five leading forecasting models demonstrate improvements in predictive performance, validating the effectiveness of our approach.

Time series forecasting is an important task that involves analyzing temporal dependencies and underlying patterns (such as trends, cyclicality, and seasonality) in historical data to predict future values or trends. Current deep learning-based forecasting models primarily employ Mean Squared Error (MSE) loss functions for regression modeling. Despite enabling direct value prediction, this method offers no uncertainty estimation and exhibits poor outlier robustness. To address these limitations, we propose OCE-TS, a novel ordinal classification approach for time series forecasting that replaces MSE with Ordinal Cross-Entropy (OCE) loss, preserving prediction order while quantifying uncertainty through probability output. Specifically, OCE-TS begins by discretizing observed values into ordered intervals and deriving their probabilities via a parametric distribution as supervision signals. Using a simple linear model, we then predict probability distributions for each timestep. The OCE loss is computed between the cumulative distributions of predicted and ground-truth probabilities, explicitly preserving ordinal relationships among forecasted values. Through theoretical analysis using influence functions, we establish that cross-entropy (CE) loss exhibits superior stability and outlier robustness compared to MSE loss. Empirically, we compared OCE-TS with five baseline models—Autoformer, DLinear, iTransformer, TimeXer, and TimeBridge—on seven public time series datasets. Using MSE and Mean Absolute Error (MAE) as evaluation metrics, the results demonstrate that OCE-TS consistently outperforms benchmark models.

Graph-level clustering (GLC), which aims to group entire graphs according to their structural and attribute-based similarities, represents a fundamental yet challenging task in various practical applications. Existing GLC methods primarily fall into two main paradigms: 1) deep graph clustering approaches based on Graph Neural Networks (GNNs), and 2) kernel-based methods that utilize predefined kernels to perform fine-grained structural comparison for clustering. However, GNN-based methods typically learn graph-level representations by aggregating node embeddings through pooling operations, which inevitably leads to substantial information loss and suboptimal clustering performance. In contrast, kernel methods, despite their theoretical expressiveness, suffer from prohibitive computational costs that hinder their scalability to large-scale settings. To solve these issues, we propose a novel graph learning framework named Anchor-driven Nyström for Deep Graph-Level Clustering (ANGC), which computes graph similarity via kernel methods while retaining the scalability of GNNs. Specifically, we first employ GNNs to encode individual graphs into sets of node embeddings. Rather than relying on pooling operations, we compute graph similarities in a kernel space constructed from these embeddings. To enhance both scalability and representational power, we introduce learnable graph Nyström anchors, which support end-to-end optimization and significantly accelerate kernel computations. To further improve the discriminative capability of these anchors, we propose the concept of anchor response discrepancy, that is, the variation in a given anchor’s responses across different samples. By maximizing this discrepancy, the anchors are encouraged to strengthen inter-graph distinctions for better clustering. Extensive experiments demonstrate the effectiveness and superiority of ANGC over existing state-of-the-art methods.

Mitigating the negative impact of noisy labels has been a perennial issue in supervised learning. Robust loss functions have emerged as a prevalent solution to this problem. In this work, we introduce the Variation Ratio as a novel property related to the robustness of loss functions, and propose a new family of robust loss functions, termed Variation-Bounded Loss (VBL), which is characterized by a bounded variation ratio. We provide theoretical analyses of the variation radio, proving that a smaller variation ratio would lead to better robustness. Furthermore, we reveal that the variation ratio provides a feasible method to relax the symmetric condition and offers a more concise path to achieve the asymmetric condition. Based on the variation ratio, we reformulate several commonly used loss functions into a variation-bounded form for pract ical applications. Positive experiments on various datasets exhibit the effectiveness and flexibility of our approach.

Large reasoning models (LRMs) have demonstrated remarkable capabilities in solving complex problems through extended chain-of-thought reasoning. However, existing approaches face a fundamental trade-off between computational efficiency and reasoning accuracy. Current methods either lack support for user-specified computational budgets or require maintaining multiple independent models, leading to significant resource overhead. In this paper, we present AdaReason, a unified framework that trains a single base model to support arbitrary user-defined computational budgets through dynamic adapter composition. Our approach introduces three key innovations: (1) a length-adaptive step reward function that stabilizes training across diverse budget constraints, (2) a progressive training strategy that gradually tightens computational bounds while maintaining model performance, and (3) a runtime adapter merging mechanism that dynamically interpolates between different computational preferences. Unlike existing methods that suffer from training instability in large context windows, AdaReason achieves stable convergence through careful reward shaping and progressive constraint tightening. Additionally, we provide a rigorous theoretical analysis, establishing a performance bound for our merged model. Experiments on different reasoning benchmarks demonstrate that AdaReason establishes a new state-of-the-art in the performance-efficiency trade-off and enables flexible runtime budget adaptation.

Recent Large Audio-Language Models (LALMs) exhibit impressive capabilities in understanding audio content for conversational QA tasks. However, these models struggle to accurately understand timestamps for temporal localization (e.g., Temporal Audio Grounding) and are restricted to short audio perception, leading to constrained capabilities on fine-grained tasks. We identify three key aspects that limit their temporal localization and long audio understanding: (i) timestamp representation, (ii) architecture, and (iii) data. To address this, we introduce TimeAudio, a novel method that empowers LALMs to connect their understanding of audio content with precise temporal perception. Specifically, we incorporate unique temporal markers to improve time-sensitive reasoning and apply an absolute time-aware encoding that explicitly grounds the acoustic features with absolute time information. Moreover, to realize end-to-end long audio understanding, we introduce a segment-level token merging module to substantially reduce audio token redundancy and enhance the efficiency of information extraction. Due to the lack of suitable datasets and evaluation metrics, we consolidate existing audio datasets into a new dataset focused on temporal tasks and establish a series of metrics to evaluate the fine-grained performance. Evaluations show strong performance across a variety of fine-grained tasks, such as dense captioning, temporal grounding, and timeline speech summarization, which demonstrates TimeAudio's robust temporal localization and reasoning capabilities.

Deep learning has shown strong performance in time series forecasting tasks. However, issues such as missing values and anomalies in sequential data hinder its further development in prediction tasks. Previous research has primarily focused on extracting feature information from sequence data or addressing these suboptimal data as positive samples for knowledge transfer. A more effective approach would be to leverage these non-ideal negative samples to enhance event prediction. In response, this study highlights the advantages of non-ideal negative samples and proposes the IdealTSF framework, which integrates both ideal positive and negative samples for time series forecasting. IdealTSF consists of three progressive steps: pretraining, training, and optimization. It first pretrains the model by extracting knowledge from negative sample data, then transforms the sequence data into ideal positive samples during training. Additionally, a negative optimization mechanism with adversarial disturbances is applied. Extensive experiments demonstrate that negative sample data unlocks significant potential within the basic attention architecture for time series forecasting. Therefore, IdealTSF is particularly well-suited for applications with noisy samples or low-quality data.

Model-based reinforcement learning (MBRL) enables efficient decision-making by learning predictive world modelsof environment dynamics. Despite recent advances, existingmodels often struggle to reconcile accurate short-term transitions with coherent long-term planning, especially in partially observable or long-horizon settings. We argue that thislimitation often stems from modeling all transitions at a single temporal resolution, which makes it challenging to simultaneously capture fine-grained local dynamics and abstractglobal structures. To this end, we propose SF-RSSM (Slow-Fast Recurrent State-Space Model), a novel method that decouples short-term and long-term dynamics via a dualbranchdesign. The fast branch captures short-horizon transitions using residual prediction, while the slow branch models long-range dependencies with a GRU-based recurrent pathway.A distillation mechanism is developed to enable cooperationacross timescales, with the slow model providing soft targetsto guide the fast model. Additionally, a curiosity module encourages exploration by promoting learning in regions wherethe fast and slow branches exhibit divergent dynamics. Experiments on CARLA, DMControl and Atari benchmarks showthat SF-RSSM outperforms strong baselines in policy performance.

Multimodal Instruction Following serves as a fundamental capability of multimodal language models, involving accurate comprehension and execution of user-provided instructions. However, existing multimodal instruction-following datasets and benchmarks face the shortcomings outlined below: (a) Lack of Difficulty Stratification, they collect diverse instruction categories but neglect the stratification of difficulty levels across these categories, which leads to overlap, bias, and low interpretability. (b) Lack of Fine-Grained Metrics, they conflate the model's ability to ``solve tasks" and ``follow constraints" into a single metric, which fails to accurately reflect its instruction-following capability. (c) Lack of Multi-Task Instructions, they overlook the fact that real-world user instructions often consist of multiple combined tasks. This paper proposes MMIFEvol, a framework for multimodal instruction evolving and benchmarking. First, we define the essential components of a carefully curated multimodal instruction set and establish corresponding difficulty levels, based on which we synthesize diverse instruction data. Next, we decouple the evaluation criteria for the instruction following into three different metrics to construct a high-quality benchmark and assess existing models. Experimental results demonstrate that current models still struggle with following complex instructions, while fine-tuning using MMIFEvol data effectively improves models' responsiveness to multimodal instructions.

Partial linear models (PLM) have attracted much attention for regression estimation and variable selection due to their feasibility on utilizing linear and nonlinear approximations jointly. However, theoretical understanding of how they control the false discovery rate (FDR) during variable selection remains limited. To address this issue, we formulate a new integral-based knockoffs (IKO) inference scheme for controlled variable selection in PLM, where integral-based knockoff statistics are used to measure the variable importance and B-splines (or random Fourier features) are employed for approximating nonlinear components. In theory, FDR control is guaranteed for both linear and nonlinear parts, and the statistical analysis for its power is established. Empirical evaluations validate the effectiveness of our proposed approach.

Large Language Models (LLMs) demonstrate strong capabilities in broad knowledge representation, yet they are inherently deficient in pixel-level perceptual understanding. Although the Segment Anything Model (SAM) represents a significant advancement in visual-prompt-driven image segmentation, it exhibits notable limitations in multi-mask prediction and category-specific segmentation tasks, and it cannot integrate all segmentation tasks within a unified model architecture. To address these limitations, we present X-SAM, a streamlined Multimodal Large Language Model (MLLM) framework that extends the segmentation paradigm from segment anything to any segmentation. Specifically, we introduce a novel unified framework that enables more advanced pixel-level perceptual comprehension for MLLMs. Furthermore, we propose a new segmentation task, termed Visual GrounDed (VGD) segmentation, which segments all instance objects with interactive visual prompts and empowers MLLMs with visual grounded, pixel-wise interpretative capabilities. To enable effective training on diverse data sources, we present a unified training strategy that supports co-training across multiple datasets. Experimental results demonstrate that X-SAM achieves state-of-the-art performance on a wide range of image segmentation benchmarks, highlighting its efficiency for multimodal, pixel-level visual understanding.

Artificial intelligence systems in critical fields like autonomous driving and medical imaging analysis often continually learn new tasks using a shared stream of input data. For instance, after learning to detect traffic signs, a model may later need to learn to classify traffic lights or different types of vehicles using the same camera feed. This scenario introduces a challenging setting we term Continual Multitask Learning (CMTL), where a model sequentially learns new tasks on an underlying data distribution without forgetting previously learned abilities. Existing continual learning methods often fail in this setting because they learn fragmented, task-specific features that interfere with one another. To address this, we introduce Learning with Preserving (LwP), a novel framework that shifts the focus from preserving task outputs to maintaining the geometric structure of the shared representation space. The core of LwP is a Dynamically Weighted Distance Preservation (DWDP) loss that prevents representation drift by regularizing the pairwise distances between latent data representations. This mechanism of preserving the underlying geometric structure allows the model to retain implicit knowledge and support diverse tasks without requiring a replay buffer, making it suitable for privacy-conscious applications. Extensive evaluations on time-series and image benchmarks show that LwP not only mitigates catastrophic forgetting but also consistently outperforms state-of-the-art baselines in CMTL tasks.

Graph classification is a critical task in analyzing graph data, with applications across various domains. While graph neural networks (GNNs) have achieved remarkable results, their ability to generalize across graphs of varying scales remains a challenge. Conventional models often perform well on large-scale graphs but struggle with distributions that are skewed towards small scales. Conversely, models tailored to address scale imbalances frequently prioritize small-scale graphs, leading to diminished performance in more balanced scenarios. To overcome these limitations, we introduce a Unbalanced-Balanced Representation Converter (U2B), which exhibits no explicit bias toward graph scales. U2B employs a two-step workflow: a distillation phase to extract base features from both node-level and graph-level representations, followed by a refinement phase to generate unbiased representations for improved balance. In the distillation phase, a static constraint guides node-level adjustments, improving the representation of nodes in small graphs. Simultaneously, a dynamic constraint in the graph-level process mitigates biases toward features from large graphs. To ensure harmony between the representations, a consistency alignment loss is introduced, aligning node-level and graph-level features to create more cohesive and balanced graph representations. Extensive experiments on multiple datasets show that U2B achieves competitive performance.

In-context learning-based medical segmentation (ICLM) enables foundation models to generalize to unseen cases without retraining. To enhance performance on test queries, existing methods typically follow a two-stage process: (1) using a retrieval encoder (RE) to map both queries and training samples into a shared feature space, and (2) retrieving and utilizing the top-k most similar training samples. While current methods fix the RE and focus on optimizing stage (2), we show that the choice of RE in stage (1) alone can account for over 70% of the performance variation, highlighting RE selection as a critical yet often overlooked factor in ICLM. In this paper, we conduct an analysis of the RE selection and make two main findings: (1) dynamically selecting the RE for each query outperforms selecting a fixed RE for the entire task; and (2) feature-space heuristics (e.g., intra-class compactness and inter-class separability) fail to predict RE quality. To this end, we propose the instance-adaptive retrieval encoder selection (IRES) method that can select the optimal RE for each query based on output predictions. IRES is based on the intuition that a good RE retrieves relevant demonstrations, helping the ICL model generate more accurate and stable segmentation masks. Thus, we introduce the shape stability score (S³), which evaluates the morphological stability of predicted masks under iterative erosion. Experiments show S³ correlates strongly with true RE quality (Pearson > 0.8), serving as a reliable selection proxy. To reduce S³’s per-query cost, we propose parallel prediction with reciprocal neighbor reuse (P2R), which accelerates inference by parallelizing encoding and reusing encoder selections across reciprocal neighbors, avoiding redundant computation. Built on S³ and P2R, IRES improves ICLM performance across FUNDUS, Brain MRI, and Chest X-ray datasets, with up to 10.6% gain on fundus segmentation.

Recent breakthroughs in Large Language Models (LLMs) have positioned them as a promising paradigm for agents, with long-term planning and decision-making emerging as core general-purpose capabilities for adapting to diverse scenarios and tasks. Real-time strategy (RTS) games serve as an ideal testbed for evaluating these two capabilities, as their inherent gameplay requires both macro-level strategic planning and micro-level tactical adaptation and action execution. Existing RTS game-based environments either suffer from relatively high computational demands or lack support for textual observations, which has constrained the use of RTS games for LLM evaluation. Motivated by this, we present TowerMind, a novel environment grounded in the tower defense (TD) subgenre of RTS games. TowerMind preserves the key evaluation strengths of RTS games for assessing LLMs, while featuring low computational demands and a multimodal observation space, including pixel-based, textual, and structured game-state representations. In addition, TowerMind supports the evaluation of model hallucination and provides a high degree of customizability. We design five benchmark levels to evaluate several widely used LLMs under different multimodal input settings. The results reveal a clear performance gap between LLMs and human experts across both capability and hallucination dimensions. The experiments further highlight key limitations in LLM behavior, such as inadequate planning validation, a lack of multifinality in decision-making, and inefficient action use. We also evaluate two classic reinforcement learning algorithms: Ape-X DQN and PPO. By offering a lightweight and multimodal design, TowerMind complements the existing RTS game-based environment landscape and introduces a new benchmark for the AI agent field.