AAAI 2026 Papers

We introduce LLA, an effective intellectual property (IP) protection scheme for generative AI models. LLA leverages the synergy between hardware and software to defend against various supply chain threats, including model theft, model corruption, and information leakage. On the software side, it embeds key bits into neurons that can trigger outliers to degrade performance and applies invariance transformations to obscure the key values. On the hardware side, it integrates a lightweight locking module into the AI accelerator while maintaining compatibility with various dataflow patterns and toolchains. An accelerator with a pre-stored secret key acts as a license to access the model services provided by the IP owner. The evaluation results show that LLA can withstand a broad range of oracle-guided key optimization attacks, while incurring a minimal computational overhead of less than 0.1% for 7,168 key bits.

Vision-language models (VLMs) have gained widespread attention for their strong zero-shot capabilities across numerous downstream tasks. However, these models assume that each test image’s class label is drawn from a predefined label set and lack a reliable mechanism to reject samples from emerging unknown classes when only unlabeled data are available. To address this gap, open-set domain adaptation methods retrain models to push potential unknowns away from known clusters. Yet, some unknown samples remain stably anchored to specific known classes in the VLM feature space due to semantic relevance, which is termed as Semantic Affinity Anchoring (SAA). Forcibly repelling these samples unavoidably distorts the native geometry of VLMs and degrades performance. Meanwhile, existing score‑based unknown detectors use simplistic thresholds and suffer from threshold sensitivity, resulting in sub‑optimal performance. To address aforementioned issues, we propose VLM-OpenXpert, which comprises two training‑free, plug‑and‑play inference modules. SUFF performs SVD on high-confidence unknowns to extract a low-rank "unknown subspace". Each sample’s projection onto this subspace is weighted and softly removed from its feature, suppressing unknown components while preserving semantics. BGAT corrects score skewness via a Box–Cox transform, then fits a bimodal Gaussian mixture to adaptively estimate the optimal threshold balancing known-class recognition and unknown-class rejection. Experiments on 9 benchmarks and three backbones (CLIP, SigLIP, ALIGN) under Source-Free OSDA settings show that our training-free pipeline matches or outperforms retraining-heavy state-of-the-art methods, establishing a powerful lightweight inference calibration paradigm for open-set VLM deployment.

Knowledge Distillation (KD) serves as an effective approach to addressing heterogeneity issues in Federated Learning (FL), leveraging additional datasets to align local and global models better. There are two primary distillation paradigms: feature-based distillation, which utilizes intermediate-layer features of the network, and logit-based distillation, which employs the final layer's logit outputs. However, existing studies often select distillation methods based on intuitive and empirical evidence when facing different heterogeneous settings, neglecting the intrinsic relationship between distillation paradigms and heterogeneity. This oversight may result in suboptimal federated knowledge distillation performance under heterogeneous conditions. In this paper, we propose the Consolidated Distillation for Heterogeneous Federated Learning - FedCD that balances knowledge representations from both feature-based and logit-based distillation to enhance performance. Specifically, to address the misalignment between knowledge conveyed by features and logits, we aggregate features from different layers via cross-layer attention to preserve semantic knowledge, followed by distribution modeling using Gaussian Mixture Models. This process strengthens knowledge distillation by constraining the transformation of different network layers' features under a consolidated distribution, thereby mitigating impacts from both data and model heterogeneity. Extensive experiments demonstrate that FedCD outperforms state-of-the-art methods by over 10.72% and validate the effectiveness of our approach.

Learning decision policies from confounded observational data is a challenging task in causal inference, as unobserved confounders can lead to biased or suboptimal actions when relying solely on machine learning models. A synergistic approach is learning to defer, which decides when to act itself and when to defer to a human expert with access to unobserved information. However, constructing the learning target, which defines the probability of choosing each action or deferral, remains a core challenge. To address this, we propose causal-target-based learning to defer (CTLD) framework, where the causal target is constructed from sharp bounds on potential outcomes. Specifically, the degree of overlap between these bounds determines the probability of deferral, while their relative positions and widths define the probabilities over actions. CTLD aligns model predictions with this causal target to make probabilistic decisions over actions and deferral. We present comprehensive theoretical guarantees for the learned policy and demonstrate the effectiveness of CTLD on synthetic and semi-synthetic datasets.

Fine-tuning large pretrained vision-language models (VLMs) has emerged as a prevalent paradigm for downstream adaptation, yet it faces a critical trade-off between domain specificity and domain generalization (DG) ability. Current methods typically fine-tune a universal model on the entire dataset, which potentially compromises the ability to generalize to unseen domains. To fill this gap, we provide a theoretical understanding of the generalization ability for VLM fine-tuning, which reveals that training multiple parameter-efficient expert models on partitioned source domains leads to better generalization than fine-tuning a universal model. Inspired by this finding, we propose a two-step domain-expert-Guided DG (GuiDG) framework. GuiDG first employs prompt tuning to obtain source domain experts, then introduces a Cross-Modal Attention module to guide the fine-tuning of the vision encoder via adaptive expert integration. To better evaluate few-shot DG, we construct ImageNet-DG from ImageNet and its variants. Extensive experiments on standard DG benchmarks and ImageNet-DG demonstrate that GuiDG improves upon state-of-the-art fine-tuning methods while maintaining efficiency.

The abundance of fine-grained spatio-temporal data, such as traffic sensor networks, offers vast opportunities for scientific discovery. However, inferring causal relationships from such observational data remains challenging, particularly due to unobserved confounders that are specific to units (e.g., geographical locations) yet influence outcomes over time. Most existing methods for spatio-temporal causal inference assume that all confounders are observed, an assumption that is often violated in practice. In this paper, we introduce Spatio-Temporal Hierarchical Causal Models (ST-HCMs), a novel graphical framework that extends hierarchical causal modeling to the spatio-temporal domain. At the core of our approach is the Spatio-Temporal Collapse Theorem, which shows that a complex ST-HCM converges to a simpler flat causal model as the amount of subunit data increases. This theoretical result enables a general procedure for causal identification, allowing ST-HCMs to recover causal effects even in the presence of unobserved, time-invariant unit-level confounders, a scenario where standard non-hierarchical models fail. We validate the effectiveness of our framework on both synthetic and real-world datasets, demonstrating its potential for robust causal inference in complex dynamic systems.

The proliferation of complex, black-box AI models has intensified the need for techniques that can explain their decisions. Feature attribution methods have become a popular solution for providing post-hoc explanations, yet the field has historically lacked a formal problem definition. This paper addresses this gap by introducing a formal definition for the problem of feature attribution, which stipulates that explanations be supported by an underlying probability distribution represented by the given dataset. Our analysis reveals that many existing model-agnostic methods fail to meet this criterion, while even those that do often possess other limitations. To overcome these challenges, we propose Distributional Feature Attribution eXplanations (DFAX), a novel, model-agnostic method for feature attribution. DFAX is the first feature attribution method to explain classifier predictions directly based on the data distribution. We show through extensive experiments that DFAX is more effective and efficient than state-of-the-art baselines.

Despite significant progress in auto-regressive image generation, inference remains slow due to the sequential nature of AR models and the ambiguity of image tokens, even when using speculative decoding. Recent works attempt to address this with relaxed speculative decoding but lack theoretical grounding. In this paper, we establish the theoretical basis of relaxed SD and propose COOL-SD, an annealed relaxation of speculative decoding built on two key insights. The first analyzes the total variation (TV) distance between the target model and relaxed speculative decoding and yields an optimal resampling distribution that minimizes an upper bound of the distance. The second uses perturbation analysis to reveal an annealing behaviour in relaxed speculative decoding, motivating our annealed design. Together, these insights enable COOL-SD to generate images faster with comparable quality, or achieve better quality at similar latency. Experiments validate the effectiveness of COOL-SD, showing consistent improvements over prior methods in speed-quality trade-offs.

Hidden degenerations in industrial time series often precede observable failures, they remain undetected by standard monitoring systems until anomalies become apparent. This gap between microscopic degradation and macroscopic observation renders conventional predictors inherently reactive, as they rely on correlations in sensor data rather than uncovering the underlying, physics‑consistent degradation states. Crucially, the microscopic mechanisms governing system evolution depend on macroscopic state variables—whose measurements are expectations over microscopic probability distributions—so purely data‑driven “top‑down” or purely physics‑guided “bottom‑up” approaches cannot forecast degeneration‑entangled industrial faults. To address these challenges, we propose a Physics-Guided Bidirectional Inference Framework that represents hidden microscopic states from macroscopic measurements. Our approach uniquely combines: (1) bottom-up physics-based simulation using Continuum Damage Mechanics to model micro-scale damage evolution under environmental stressors, and (2) top-down probabilistic inference via maximum entropy formalism to estimate latent microstate distributions from sparse sensor data. This bidirectional mechanism enables early failure prediction by bridging observable measurements with unobservable degeneration. Validation on real-world railway infrastruc datasets demonstrates significant improvements in early fault prediction compared to state-of-the-art baselines. Our method establishes a new paradigm for safety-critical industrial applications requiring reliable prediction of hidden degeneration processes.

Semi-Supervised Learning (SSL) aims to improve the learning performance of supervised learning with a large number of unlabeled samples. The existing SSL methods such as FixMatch and FlexMatch select unlabeled samples with high-confident pseudo-labels and make consistency constraints between their weak and strong augmentations. Unfortunately, they cannot be applied Semi-Supervised Regression (SSR) because regression predictions can not reflect the confidence of pseudo-labels. To solve this, a recent SSR method RankUp incorporates an auxiliary ranking task by leveraging sample pairs with high-confident pseudo-ranks. In this paper, we upgrade Rankup to a novel SSR method, namely Semi-Supervised Regression by Ranking Close Unlabeled Samples (SSR-RCUS). Its basic idea is reconstructing closed mixup augmented samples with high-confident pseudo-ranks under a monotonicity assumption, and then applying them to the auxiliary ranking task to improve regression performance. We conduct extensive experiments to evaluate the performance of SSR-RCUS on benchmark datasets, and empirical results demonstrate that SSR-RCUS can outperform the existing baselines in various settings, especially when labeled data are scarce.

Cross-Domain Few-Shot Learning (CD-FSL) remains a significant challenge due to substantial distribution shifts between source and target domains. While prior approaches primarily focus on spatial alignment, they often overlook discrepancies in the frequency domain. In this paper, we reveal frequency band discretization as a key phenomenon, characterized by intra-domain low-frequency dominance, inter-domain amplitude divergence, and limited high-frequency variation. This spectral disharmony biases models toward low-frequency components, leading to spectral collapse. We quantify spectral collapse via the effective rank, a principled measure of spectral diversity. To mitigate spectral collapse, we propose Harmonized Amplitude Perturbation (HAP), a frequency-domain augmentation strategy that perturbs the amplitude spectrum via frequency-aware gains sampled from Harmonized Distributions, while fixing the phase spectrum to maintain semantic integrity. Extensive experiments on both Cross-Domain Few-Shot Image Classification and Object Detection benchmarks demonstrate that HAP effectively increases spectral diversity and consistently improves generalization, outperforming state-of-the-art methods without introducing extra model complexity.

Despite significant advancements in general AI, its effectiveness in the medical domain is limited by the lack of specialized medical knowledge. To address this, we formulate GMAI-VL-5.5M, a multimodal medical dataset created by converting hundreds of specialized medical datasets with various annotations into high-quality image-text pairs. This dataset offers comprehensive task coverage, diverse modalities, and rich image-text data. Building upon this dataset, we develop GMAI-VL, a 7B-parameter general medical vision-language model, with a three-stage training strategy that enhances the integration of visual and textual information. This approach significantly improves the model's ability to process multimodal data, supporting accurate diagnoses and clinical decision-making. Experiments show that GMAI-VL achieves state-of-the-art performance across various multimodal medical tasks, including visual question answering and medical image diagnosis.

In offline-to-online (O2O) reinforcement learning, achieving efficient performance improvement while maintaining training stability remains a critical challenge for effective fine-tuning. Existing O2O methods usually focus on the balance between policy improvement and policy constraint during online fine-tuning. However, they often overlook sample differences, leading to suboptimal performance. To address this challenge, we identify that the effectiveness of policy learning exhibits significant variation across states. Therefore, we propose the notion of state proficiency to capture the degree of effective learning in a given state. We propose State Proficiency-Based Adaptive Fine-Tuning (SPA), a straightforward yet effective method that establishes proficiency-based sample priorities in policy optimization to facilitate effective fine-tuning. Specifically, SPA focuses on low proficiency samples during policy improvement to enhance sample efficiency, while emphasizing high proficiency samples during policy constraint to ensure stable training. Extensive empirical results demonstrate that SPA achieves significant improvements over existing methods, attaining state-of-the-art performance on the D4RL benchmark.

Text-to-MIDI generation offers editable and hierarchical control over symbolic music generation. Previous approaches either convert text into a limited set of musical attributes and generate music based on these attributes, which limits semantic controllability, or use end-to-end models that map text directly to music without deeply aligning the features of both modalities, often resulting in a lack of structural coherence and mismatches in key, meter, and tempo. We propose MIDILM, which addresses these limitations by employing text conditioning with a dual-path decoder that processes textual and musical information through separate feedforward paths following a shared masked self-attention mechanism. On the MidiCaps benchmark, MIDILM outperformed the strongest baseline, with relative improvements ranging from 6.07% on CLAP to 144.77% on TB across semantic alignment and structural metrics. These gains confirm its ability to enhance both semantic controllability and structural coherence. Collectively, we expect that MIDILM will serve as a useful reference framework for future investigations into controllable and structurally faithful cross-modal music generation.

Designing molecules with desired properties, aka the oRiented molEcule Design (RED), is a fundamental task in chemistry and materials science. While graph diffusion models (GDMs) and reinforcement learning techniques (RL) show promise in molecule structure generation and property optimization stages individually, their integration in the unified RED task often suffers from poor compatibility. The large variance among candidate molecular structures generated by GDMs can be amplified in the iterative optimization process of RL, leading to slow and unstable convergence. In this work, motivated by the adaptive and divide-and-conquer characteristics of Mixture of Experts (MoE) architecture, we propose a novel framework called MoE-Guided Graph Diffusion Model (MEGD) that incorporates the MoE architecture to guide the orchestration of GDM and RL, promoting faster and more stable convergence in the design process. MEGD is evaluated on benchmark datasets optimizing the physical and chemical properties of AI-generated molecular structures. On all three datasets, our method outperforms the best of 9 alternative models by 7.73% on the target structural properties, while not penalizing other important application-level quality metrics of the generated molecules. A real-world case study on an emerging class of material, i.e., metal-organic framework, is also conducted, which further demonstrates the effectiveness of our method in accomplishing the RED task.

Multi-agent systems (MAS) based on large language models (LLMs) have emerged as a powerful solution for dealing with complex problems across diverse domains. The effectiveness of MAS is critically dependent on its collaboration topology, which has become a focal point for automated design research. However, existing approaches are fundamentally constrained by their reliance on a template graph modification paradigm with a predefined set of agents and hard-coded interaction structures, significantly limiting their adaptability to task-specific requirements. To address these limitations, we reframe MAS design as a conditional autoregressive graph generation task, where both the system composition and structure are designed jointly. We propose ARG-Designer, a novel autoregressive model that operationalizes this paradigm by constructing the collaboration graph from scratch. Conditioned on a natural language task query, ARG-Designer sequentially and dynamically determines the required number of agents, selects their appropriate roles from an extensible pool, and establishes the optimal communication links between them. This generative approach creates a customized topology in a flexible and extensible manner, precisely tailored to the unique demands of different tasks. Extensive experiments across six diverse benchmarks demonstrate that ARG-Designer not only achieves state-of-the-art performance but also enjoys significantly greater token efficiency and enhanced extensibility.

As the pretraining-finetuning paradigm becomes dominant in modern AI, the security of model supply chains faces new risks from backdoor attacks. Existing work primarily studies backdoors injected during pretraining and treats subsequent finetuning with clean data as a defense, while recent finetuning-activated attacks assume white-box access to the downstream data distribution, which is rarely realistic in practice. We introduce Dormant Backdoor, a finetuning-activated attack that requires no prior knowledge of downstream tasks. Instead of binding the backdoor to static input patterns, Dormant Backdoor exploits the universal dynamics of gradient-based optimization as a process-as-trigger mechanism. We formulate the attack as a bilevel optimization problem that simulates the victim's finetuning trajectory on proxy data, and jointly optimizes the poisoned model and trigger under lethality, utility, and stealth objectives. Before finetuning, the poisoned model remains behaviorally close to a clean model and can evade existing backdoor detectors; after finetuning, the same adaptation process reliably amplifies the backdoor on diverse downstream datasets and finetuning strategies. Our results reveal a previously underexplored class of process-as-trigger vulnerabilities and highlight the need for defenses that explicitly secure the model adaptation process.

Reinforcement learning (RL) has demonstrated considerable potential for enhancing reasoning in large language models (LLMs). However, existing methods suffer from Gradient Starvation and Policy Degradation when training directly on samples with mixed difficulty. To mitigate this, prior approaches leverage Chain-of-Thought (CoT) data, but the construction of high-quality CoT annotations remains labor-intensive. Alternatively, curriculum learning strategies have been explored but frequently encounter challenges, such as difficulty mismatch, reliance on manual curriculum design, and catastrophic forgetting. To address these issues, we propose AdaCuRL, a Adaptive Curriculum Reinforcement Learning framework that integrates coarse-to-fine difficulty estimation with adaptive curriculum scheduling. This approach dynamically aligns data difficulty with model capability and incorporates a data revisitation mechanism to mitigate catastrophic forgetting. Furthermore, AdaCuRL employs adaptive reference and sparse KL strategies to prevent Policy Degradation. Extensive experiments across diverse reasoning benchmarks demonstrate that AdaCuRL consistently achieves significant performance improvements on both LLMs and MLLMs.

Pseudo-random number generators (PRNGs) are high-nonlinear processes, and they are key blocks in optimization of Large language models. Transformers excel at processing complex nonlinear relationships. Thus it is reasonable to generate high-quality pseudo-random numbers based on transformers. In this paper, we explore this question from both theoretical and practical perspectives, highlighting the potential benefits and implications of Transformer in PRNGs. We theoretically demonstrate that decoder-only Transformer models with Chain-of-Thought can simulate both the Linear Congruential Generator (LCG) and Mersenne Twister (MT) PRNGs. Based on this, we conclude that the log-precision decoder-only Transformer can represent non-uniform AC0. Our simulative theoretical findings are validated through experiments. The random numbers generated by Transformer-based PRNGs successfully pass the majority of NIST tests, whose heat maps exhibit clear statistical randomness. Finally, we assess their capability in prediction attacks.