AAAI 2026 Papers

Recent research reveals that a minority of high-entropy tokens significantly influence the reasoning quality of large language models (LLMs). Inspired by this, we propose Prototype Entropy Alignment (PEA), a reinforcement learning framework that models effective reasoning not as a single path but as a collection of learnable "entropy signatures." PEA identifies these signatures by clustering expert trajectories' uncertainty patterns into a diverse and continuously updated set of prototypes. The model is then rewarded for aligning its own reasoning process with these evolving targets, creating a self-improvement loop. Instead of replacing traditional outcome-based rewards, PEA provides a complementary, process-oriented signal. Our experiments show that this synergy is crucial: PEA substantially boosts performance on creative and general reasoning tasks and, when combined with outcome rewards, achieves SOTA results on structured tasks such as mathematics. By rewarding alignment with diverse and evolving reasoning structures, PEA offers a robust, verifier-free pathway to enhance reasoning's adaptability.

Federated Edge Learning (FEL) has emerged as a promising approach for enabling edge devices to collaboratively train machine learning models while preserving data privacy. Despite its advantages, practical FEL deployment faces significant challenges related to device constraints and device-server interactions, necessitating heterogeneous, user-adaptive model training with limited and uncertain communication. While knowledge cache-driven federated learning offers a promising FEL solution for demanding edge environments, its logits-based interaction design provides poor richness of exchanged information for on-device model optimization. To tackle this issue, we introduce DistilCacheFL, a novel personalized FEL architecture that enhances the exchange of optimization insights while delivering state-of-the-art performance with efficient communication. DistilCacheFL incorporates the benefits of both dataset distillation and knowledge cache-driven federated learning by storing and organizing distilled data as knowledge in the server-side knowledge cache, allowing devices to periodically download and utilize personalized knowledge for local model optimization. Moreover, a device-centric cache sampling strategy is introduced to tailor transferred knowledge for individual devices within controlled communication bandwidth. Extensive experiments on five datasets covering image recognition, audio understanding, and mobile sensor data mining tasks demonstrate that (1) DistilCacheFL significantly outperforms state-of-the-art methods regardless of model structures, data distributions, and modalities. (2) DistilCacheFL can train splendid personalized on-device models with at least 28.6 improvement in communication efficiency.

Graph anomaly detection (GAD), which aims to detect outliers in graph-structured data, has received increasing research attention recently. However, existing GAD methods assume identical training and testing distributions, which is rarely valid in practice. In real-world scenarios, unseen but normal samples may emerge during deployment, leading to a normality shift that degrades the performance of GAD models trained on the original data. Through empirical analysis, we reveal that the degradation arises from (1) semantic confusion, where unseen normal samples are misinterpreted as anomalies due to their novel patterns, and (2) aggregation contamination, where the representations of seen normal nodes are distorted by unseen normals through message aggregation. While retraining or fine-tuning GAD models could be a potential solution to the above challenges, the high cost of model retraining and the difficulty of obtaining labeled data often render this approach impractical in real-world applications. To bridge the gap, we proposed a lightweight and plug-and-play Test-time adaptation framework for correcting Unseen Normal pattErns (TUNE) in GAD. To address semantic confusion, a graph aligner is employed to align the shifted data to the original one at the graph attribute level. Moreover, we utilize the minimization of representation-level shift as a supervision signal to train the aligner, which leverages the estimated aggregation contamination as a key indicator of normality shift. Extensive experiments on 10 real-world datasets demonstrate that TUNE significantly enhances the generalizability of pre-trained GAD models to both synthetic and real unseen normal patterns.

Asynchronous Federated Learning (AFL) is acclaimed for accelerating collaborative training on heterogeneous systems by eliminating the wait for stragglers. While current solutions focus on improving convergence amidst update delays, they neglect how delayed aggregation fosters free-riding attacks, allowing malicious clients to easily extract the global model without contribution. This behavior results in significant fairness issues and performance degradation. To address this challenge, we propose OPTION, the first online pricing strategy tailored to mitigate free-riding in AFL. OPTION establishes an economic model in which access to model updates is purchased using credits earned from verified contributions. Specifically, OPTION values each model update according to its marginal performance gain and training cost, and subsequently necessitates a download fee from each client based on the Hotelling model to prevent zero-cost acquisition. Moreover, OPTION rewards clients for successful updates under non-arbitrage constraints, effectively balancing individual utility and task budget. To maximize the average model performance while satisfying these conditions, OPTION leverages the Lyapunov drift framework and a probabilistic sampling-based algorithm to optimize the pricing parameters. Extensive experimental results on three real-world datasets demonstrate that OPTION effectively mitigates freeriding attacks in AFL, increases the number of valid updates by at least 23.97%, and achieves a model accuracy improvement of at least 3.01% compared to state-of-the-art baselines.

We consider the problem of risk-sensitive control in a reinforcement learning (RL) framework. In particular, we aim to find a risk-optimal policy by maximizing the distortion riskmetric (DRM) of the discounted reward in a finite-horizon Markov decision process (MDP). DRMs are a rich class of risk measures that include several well-known risk measures as special cases. We derive a policy Hessian theorem for the DRM objective using the likelihood ratio method. Using this result, we propose a natural DRM Hessian estimator from sample trajectories of the underlying MDP. Next, we present a cubic-regularized policy Newton algorithm for solving this problem in an on-policy RL setting using estimates of the DRM gradient and Hessian. Our proposed algorithm is shown to converge to an ϵ-second-order stationary point (ϵ-SOSP) of the DRM objective, and this guarantee ensures the escaping of saddle points. The sample complexity of our algorithms to find an ϵ-SOSP is O(ϵ−3.5). Our experiments validate the theoretical findings. To the best of our knowledge, our is the first work to present convergence to an ϵ-SOSP of a risk-sensitive objective, while existing works in the literature have either shown convergence to a first-order stationary point of a risk-sensitive objective, or a SOSP of a risk-neutral one.

Evolutionary algorithms (EAs) are optimization algorithms that simulate natural selection and genetic mechanisms. Despite advancements, existing EAs have two main issues: (1) they rarely update next-generation individuals based on global correlations, thus limiting comprehensive learning; (2) it is challenging to balance exploration and exploitation, excessive exploitation leads to premature convergence to local optima, while excessive exploration results in an excessively slow search. Existing EAs heavily rely on manual parameter settings, inappropriate parameters might disrupt the exploration-exploitation balance, further impairing model performance. To address these challenges, we propose a novel evolutionary algorithm framework called Graph Neural Evolution (GNE). Unlike traditional EAs, GNE represents the population as a graph, where nodes correspond to individuals, and edges capture their relationships, thus effectively leveraging global information. Meanwhile, GNE utilizes spectral graph neural networks (GNNs) to decompose evolutionary signals into their frequency components and designs a filtering function to fuse these components. High-frequency components capture diverse global information, while low-frequency components capture more consistent information. This explicit frequency filtering strategy directly controls global-scale features through frequency components, overcoming the limitations of manual parameter settings and making the exploration-exploitation control more interpretable and effective. Extensive evaluations on nine benchmark functions (e.g., Sphere, Rastrigin, and Rosenbrock) demonstrate that GNE consistently outperforms both classical algorithms (GA, DE, CMA-ES) and advanced algorithms (SDAES, RL-SHADE) under various conditions, including original, noise-corrupted, and optimal solution deviation scenarios. GNE achieves solution quality several orders of magnitude better than other algorithms (e.g., 3.07e-20 mean on Sphere vs. 1.51e-07).

Recent self-supervised pre-training methods for object detection often rely on generic object proposals for localization and semantic feature learning for classification, but they yield limited improvements when applied to Detection Transformers (DETR) due to a lack of architectural alignment. Hence, we propose an elegant and versatile self-supervised framework tailored for DETR-like models called Distance-aware Multi-view Contrastive Learning (DisCo DETR). DisCo DETR enhances localization and semantic features through two core components. (i) Distance-aware Multi-view Object Query Fusion explicitly guides object queries to focus on spatially close objects across views, stabilizing training and improving localization accuracy. (ii) Contrastive Learning for DETR uses native bipartite matching to identify positive output pairs across views and pull them closer, enhancing semantic features discrimination with no extra matching. DisCo DETR can be seamlessly integrated into DETR-like models and achieves SOTA transfer performance on PASCAL VOC and COCO benchmarks across multiple variants.

Dynamic sequences with varying lengths have been widely used in the training of Transformer-based large language models (LLMs). However, current training frameworks adopt a pre-defined static parallel strategy for these sequences, causing neither communication-parallelization cancellation on short sequences nor out-of-memory on long sequences. To mitigate these issues, we propose ParaDySe, a novel adaptive Parallel strategy switching framework for Dynamic Sequences. ParaDySe enables on-the-fly optimal strategy adoption according to the immediate input sequence. It first implements the modular function libraries for parallel strategies with unified tensor layout specifications, and then builds sequence-aware memory and time cost models with hybrid methods. Guided by cost models, ParaDySe selects optimal layer-wise strategies for dynamic sequences via an efficient heuristic algorithm. By integrating these techniques together, ParaDySe achieves seamless hot-switching of optimal strategies through its well-designed function libraries. We compare ParaDySe with baselines on representative LLMs under datasets with sequence lengths up to 624K. Experimental results indicate that ParaDySe addresses OOM and CPC bottlenecks in LLM training by systematically integrating long-sequence optimizations with existing frameworks.

The strong lottery ticket hypothesis (SLTH) conjectures that high-performing subnetworks, called strong lottery tickets (SLTs), are hidden in randomly initialized neural networks. Although recent theoretical studies have established the SLTH across various neural architectures, the SLTH for transformer architectures still lacks theoretical understanding. In particular, the current theory of the SLTH does not yet account for the multi-head attention (MHA) mechanism, a core component of transformers. To address this gap, we introduce a theoretical analysis of the existence of SLTs within MHAs. We prove that, if a randomly initialized MHA of H heads and input dimension d has the hidden dimension O(d log(Hd^(3/2))) for the key and value, it contains an SLT that approximates an arbitrary MHA with the same input dimension with high probability. Furthermore, by leveraging this theory for MHAs, we extend the SLTH to transformers without normalization layers. We empirically validate our theoretical findings, demonstrating that the approximation error between the SLT within a source model (MHA and transformer) and an approximate target counterpart decreases exponentially by increasing the hidden dimension of the source model.

Foundation models have revolutionized artificial intelligence across numerous domains, yet their transformative potential remains largely untapped in Extreme Multi-label Classification (XMC). Queries in XMC are associated with relevant labels from extremely large label spaces, where it is critical to strike a balance between efficiency and performance. Therefore, many recent approaches efficiently pose XMC as a maximum inner product search between embeddings learned from small encoder-only transformer architectures. In this paper, we address two important aspects in XMC: how to effectively harness larger decoder-only models, and how to exploit visual information while maintaining computational efficiency. We demonstrate that both play a critical role in XMC separately and can be combined for improved performance. We show that a few billion-size decoder can deliver substantial improvements while keeping computational overhead manageable. Furthermore, our Vision-enhanced eXtreme Multi-label Learning framework (ViXML) efficiently integrates foundation vision models by pooling a single embedding per image. This limits computational growth while unlocking multi-modal capabilities. Remarkably, ViXML with small encoders outperforms text-only decoder in most cases, showing that an image is worth billions of parameters. Finally, we present an extension of existing text-only datasets to exploit visual metadata and make them available for future benchmarking. Comprehensive experiments across four public text-only datasets and their corresponding image enhanced versions validate our proposals' effectiveness, surpassing previous state-of-the-art by up to +8.21% in P@1 on the largest dataset.

Machine learning models now drive many critical decisions, making explanations of their reasoning essential. Recent work analyzes the complexity of exact explanations in transparent models, but these explanations are often too large for practical use. This has motivated research into probabilistic alternatives. We study probabilistic extensions that allow controlled uncertainty while maintaining rigorous foundations. We analyze three basic model types: decision trees, decision lists, and decision sets. We introduce algorithms for computing both local and global probabilistic explanations for these models. Our main result shows that computing minimum-size probabilistic explanations is fixed-parameter tractable when parameterized by structural properties---specifically, the number of terms for decision lists and decision sets and the minimum of the number of positive and the number of negative leaves.

In this work, we propose a novel framework for the logical specification of non-Markovian rewards in Markov Decision Processes (MDPs) with large state spaces. Our approach leverages Linear Temporal Logic Modulo Theories over finite traces (LTLfMT), a more expressive extension of classical temporal logic in which predicates are first-order formulas of arbitrary first-order theories rather than simple Boolean variables. This enhanced expressiveness enables the specification of complex tasks over unstructured and heterogeneous data domains, promoting a unified and reusable framework that eliminates the need for manual predicate encoding. However, the increased expressive power of LTLfMT introduces additional theoretical and computational challenges compared to standard LTLf specifications. We address these challenges from a theoretical standpoint, identifying a fragment of LTLfMT that is tractable but sufficiently expressive for reward specification in an infinite-state-space context. From a practical perspective, we introduce a method based on reward machines and Hindsight Experience Replay (HER) to translate first-order logic specifications and address reward sparsity. We evaluate this approach to a continuous-control setting using Non-Linear Arithmetic Theory, showing that it enables natural specification of complex tasks. Experimental results show how a tailored implementation of HER is fundamental in solving tasks with complex goals.

Multi-task learning (MTL) is critical in real-world applications such as autonomous driving and robotics, enabling simultaneous handling of diverse tasks. However, obtaining fully annotated data for all tasks is impractical due to labeling costs. Existing methods for partially labeled MTL typically rely on predictions from unlabeled tasks, making it difficult to establish reliable task associations and potentially leading to negative transfer and suboptimal performance. To address these issues, we propose a prototype-based knowledge retrieval framework that achieves robust MTL instead of relying on predictions from unlabeled tasks. Our framework consists of two key components: (1) a task prototype embedding task-specific characteristics and quantifying task associations, and (2) a knowledge retrieval transformer that adaptively refines feature representations based on these associations. To achieve this, we introduce an association knowledge generating (AKG) loss to ensure the task prototype consistently captures task-specific characteristics. Extensive experiments demonstrate the effectiveness of our framework, highlighting its potential for robust multi-task learning, even when only a subset of tasks is annotated.

Modeling continuous-time dynamics from sparse and irregularly-sampled time series remains a fundamental challenge. Neural controlled differential equations provide a principled framework for such tasks, yet their performance is highly sensitive to the choice of control path constructed from discrete observations. Existing methods commonly employ fixed interpolation schemes, which impose simplistic geometric assumptions that often misrepresent the underlying data manifold, particularly under high missingness. We propose FlowPath, a novel approach that learns the geometry of the control path via an invertible neural flow. Rather than merely connecting observations, FlowPath constructs a continuous and data-adaptive manifold, guided by invertibility constraints that enforce information-preserving and well-behaved transformations. This inductive bias distinguishes FlowPath from prior unconstrained learnable path models. Empirical evaluations on 18 benchmark datasets and a real-world case study demonstrate that FlowPath consistently achieves statistically significant improvements in classification accuracy over baselines using fixed interpolants or non-invertible architectures. These results highlight the importance of modeling not only the dynamics along the path but also the geometry of the path itself, offering a robust and generalizable solution for learning from irregular time series.

World models have been developed to support sample-efficient deep reinforcement learning agents. However, it remains challenging for world models to accurately replicate environments that are high-dimensional, non-stationary, and composed of multiple objects with rich interactions since most world models learn holistic representations of all environmental components. By contrast, humans perceive the environment by decomposing it into discrete objects, facilitating efficient decision-making. Motivated by this insight, we propose Slot Transformer Imagination with CAusality-aware reinforcement learning (STICA), a unified framework in which object-centric Transformers serve as the world model and causality-aware policy and value networks. STICA represents each observation as a set of object-centric tokens, together with tokens for the agent action and the resulting reward, enabling the world model to predict token-level dynamics and interactions. The policy and value networks then estimate token-level cause--effect relations and use them in the attention layers, yielding causality-guided decision-making. Experiments on object-rich benchmarks demonstrate that STICA consistently outperforms state-of-the-art agents in both sample efficiency and final performance.

Pretrained vision-language models (VLMs), especially CLIP, excel at adapting to downstream tasks through fine-tuning with sufficient high-quality labeled data. However, real-world training data often contains noisy labels, leading to significant performance degradation when models are naively fine-tuned on them. Existing noisy label learning methods for VLMs typically leverage the model's own pretrained knowledge, either via zero-shot predictions or vanilla self-training based on them, to identify and handle noisy samples. Crucially, these approaches blindly trust the VLM's pretrained knowledge, which can introduce endogenous confirmation bias: erroneous pretrained priors lead to incorrect noise detection, further amplifying the bias and corrupting the model. To overcome this limitation, we propose the Debiased Knowledge Adaptation Framework (DKAF), which empowers the model to challenge and correct potentially flawed zero-shot predictions. DKAF operates in three progressive phases: (1) Clean Sample Selection. We introduce a cross-modal collaborative pseudo-labeling to train a robust noisy label detector, explicitly mitigating confirmation bias by aggregating diverse signals beyond the model's initial zero-shot view. (2) Noisy Label Refinement. For samples identified as noisy, we apply a dual-modal consistency strategy to selectively correct their labels, leveraging alignment between dominant and fused modalities to guide refinement while minimizing reliance on potentially biased internal knowledge. (3) Model Adaptation. The model is progressively fine-tuned using the jointly curated dataset of selected clean samples and corrected noisy samples, promoting robust adaptation to the target task. Extensive experiments on nine benchmark datasets (both synthetic and real-world noise) demonstrate that DKAF consistently outperforms state-of-the-art multimodal noisy label learning methods. Notably, under high-noise conditions, DKAF achieves average accuracy improvements of 3.08%.

Hierarchical representations provide powerful and principled approaches for analyzing many musical genres. Such representations have been broadly studied in music theory, for instance via Schenkerian analysis (SchA). Hierarchical music analyses, however, are highly cost-intensive; the analysis of a single piece of music requires a great deal of time and effort from trained experts. The representation of hierarchical analyses in a computer-readable format is also a further challenge. Given recent developments in hierarchical deep learning and increasing quantities of computer-readable data, there is great promise in extending such work for an automatic hierarchical representation framework. This paper thus introduces a novel approach, AutoSchA, which extends recent developments in graph neural networks (GNNs) for hierarchical music analysis. AutoSchA features three key contributions: 1) a new graph learning framework for hierarchical music representation, 2) a new graph pooling mechanism based on node isolation that directly optimizes learned pooling assignments, and 3) a state-of-the-art architecture that integrates such developments for automatic hierarchical music analysis. We show, in a suite of experiments, that AutoSchA performs comparably to human experts when analyzing Baroque fugue subjects.

Graph-Level Anomaly Detection (GLAD) seeks to identify anomalous graphs within graph datasets, which has significant applications across diverse real-world fields. Most existing GLAD methods are trained in an unsupervised manner due to high costs for labeling, resulting in sub-optimal performance when compared to supervised methods. To fill this gap, we propose a Disentangled Generation-Based Prototypical Alignment (DGPA) method that extends graph-level anomaly detection to Few-Shot Unsupervised Domain Adaptation (FUDA) setting, aiming to identify anomalous graphs from a set of unlabeled graphs (target domain) by using partially labeled graphs from a different but related domain (source domain), which fulfills the practical requirement of transferring anomaly knowledge. This is specifically achieved through a dedicated Disentangled Sample Generation module, which addresses label scarcity by generating faithful samples with disentangled representation learning grounded in Information Bottleneck principle, along with a Graph-based Prototypical Self-Supervision module, which alleviates domain shift by encoding and aligning semantic structures in the shared latent space across domains in a self-supervised manner. Extensive experiments on five benchmark datasets reveal the effectiveness of our proposed DGPA.

We introduce PKR-QA (Procedural Knowledge Reasoning Question Answering), a new benchmark for question answering over procedural tasks that require structured reasoning. PKR-QA is constructed semi-automatically using a procedural knowledge graph (PKG), which encodes task-specific knowledge across diverse domains. The PKG is built by curating and linking information from the COIN instructional video dataset and the ontology, enriched with commonsense knowledge from ConceptNet and structured outputs from Large Language Models (LLMs), followed by manual verification. To generate question-answer pairs, we design graph traversal templates where each template is applied systematically over PKG. To enable interpretable reasoning, we propose a neurosymbolic approach called Knowledge Module Learning (KML), which learns procedural relations via neural modules and composes them for structured reasoning with LLMs. Experiments demonstrate that this paradigm improves reasoning performance on PKR-QA and enables step-by-step reasoning traces that facilitate interpretability.