AAAI 2025 论文列表

Offline preference-based reinforcement learning (PbRL) typically operates in two phases: first, use human preferences to learn a reward model and annotate rewards for a reward-free offline dataset; second, learn a policy by optimizing the learned reward via offline RL. However, accurately modeling step-wise rewards from trajectory-level preference feedback presents inherent challenges. The reward bias introduced, particularly the overestimation of predicted rewards, leads to optimistic trajectory stitching, which undermines the pessimism mechanism critical to the offline RL phase. To address this challenge, we propose In-Dataset Trajectory Return Regularization (DTR) for offline PbRL, which leverages conditional sequence modeling to mitigate the risk of learning inaccurate trajectory stitching under reward bias. Specifically, DTR employs Decision Transformer and TD-Learning to strike a balance between maintaining fidelity to the behavior policy with high in-dataset trajectory returns and selecting optimal actions based on high reward labels. Additionally, we introduce an ensemble normalization technique that effectively integrates multiple reward models, balancing the trade-off between reward differentiation and accuracy. Empirical evaluations on various benchmarks demonstrate the superiority of DTR over other state-of-the-art baselines.

In this work, we present an in-context policy adaptation (ICPAD) framework designed for long-horizon multi-task environments, exploring diffusion-based skill learning techniques in cross-domain settings. The framework enables rapid adaptation of skill-based reinforcement learning policies to diverse target domains, especially under stringent constraints on no model updates and only limited target domain data. Specifically, the framework employs a cross-domain skill diffusion scheme, where domain-agnostic prototype skills and a domain-grounded skill adapter are learned jointly and effectively from an offline dataset through cross-domain consistent diffusion processes. The prototype skills act as primitives for common behavior representations of long-horizon policies, serving as a lingua franca to bridge different domains. Furthermore, to enhance the in-context adaptation performance, we develop a dynamic domain prompting scheme that guides the diffusion-based skill adapter toward better alignment with the target domain. Through experiments with robotic manipulation in Metaworld and autonomous driving in CARLA, we show that our ICPAD framework achieves superior policy adaptation performance under limited target domain data conditions for various cross-domain configurations including differences in environment dynamics, agent embodiment, and task horizon.

Decision trees, owing to their interpretability, are attractive as control policies for (dynamical) systems. Unfortunately, constructing, or synthesising, such policies is a challenging task. Previous approaches do so by imitating a neural-network policy, approximating a tabular policy obtained via formal synthesis, employing reinforcement learning, or modelling the problem as a mixed-integer linear program. However, these works may require access to a hard-to-obtain accurate policy or a formal model of the environment (within reach of formal synthesis), and may not provide guarantees on the quality or size of the final tree policy. In contrast, we present an approach to synthesise optimal decision-tree policies given a deterministic black-box environment and specification, a discretisation of the tree predicates, and an initial set of states, where optimality is defined with respect to the number of steps to achieve the goal. Our approach is a specialised search algorithm which systematically explores the (exponentially large) space of decision trees under the given discretisation. The key component is a novel trace-based pruning mechanism that significantly reduces the search space. Our approach represents a conceptually novel way of synthesising small decision-tree policies with optimality guarantees even for black-box environments with black-box specifications.

The MaxSAT problem is an optimization version of the satisfiability problem (SAT). A tight lower bound (LB) on the number of falsified soft clauses in a MaxSAT solution is crucial for the efficiency of Branch-and-Bound (BnB) MaxSAT solvers. To compute an LB, modern BnB solvers detect disjoint inconsistent subsets of soft clauses, called cores, using unit propagation. A notable feature of these solvers is that soft clauses belonging to already detected cores cannot be reused to detect additional cores, limiting the number of cores that can be detected. In this paper, we propose an unlocking mechanism that allows the reuse of soft clauses in already detected cores while ensuring the soundness of LB. Experimental results show that this unlocking mechanism consistently improves the performance of a state-of-the-art BnB solver. In addition, it allowed us to win the first two places in the exact unweighted category of the MaxSAT Evaluation 2024.

Classical Transformer-based line segment detection methods have delivered impressive results. However, we observe that some accurately detected line segments are assigned low confidence scores during prediction, causing them to be ranked lower and potentially suppressed. Additionally, these models often require prolonged training periods to achieve strong performance, largely due to the necessity of bipartite matching. In this paper, we introduce RANK-LETR, a novel Transformer-based line segment detection method. Our approach leverages learnable geometric information to refine the ranking of predicted line segments by enhancing the confidence scores of high-quality predictions in a posterior verification step. We also propose a new line segment proposal method, wherein the feature point nearest to the centroid of the line segment directly predicts the location, significantly improving training efficiency and stability. Moreover, we introduce a line segment ranking loss to stabilize rankings during training, thereby enhancing the generalization capability of the model. Experimental results demonstrate that our method outperforms other Transformer-based and CNN-based approaches in prediction accuracy while requiring fewer training epochs than previous Transformer-based models.

The Retrieval-Augmented Language Model (RALM) has demonstrated remarkable performance on knowledge-intensive tasks by integrating external knowledge during inference, which mitigates the factual hallucinations inherited in large language models (LLMs). Despite these advancements, challenges persist in the implementation of RALMs, particularly in terms of reliability and traceability. Specifically, the irrelevant document retrieval may result in unhelpful responses or even deteriorate the performance of LLMs, while the lack of appropriate citations in outputs complicates efforts to verify the trustworthiness of the models. To this end, we propose a novel self-reasoning framework aimed at improving the reliability and traceability of RALMs, whose core idea is to leverage reasoning trajectories generated by the LLM itself. The framework involves constructing self-reasoning trajectories through three processes: a relevance-aware process, an evidence-aware selective process, and a trajectory analysis process. We evaluated our framework across four public datasets (two short-form QA datasets, one long-form QA dataset, and one fact verification dataset) to demonstrate its superiority. Our method can outperform existing state-of-the-art models and achieve performance comparable with GPT-4, using only 2,000 training samples.

We introduce a novel matrix representation for differentially private training and prediction methods tailored to random forest classifiers. Our approach involves representing each root-to-leaf decision path in all trees as a row vector in a matrix. Similarly, inference queries are represented as a matrix. This representation enables us to collectively analyze privacy across multiple trees and inference queries, resulting in optimal DP noise allocation under the Laplace Mechanism. Our experimental results show significant accuracy improvements of up to 40% compared to state-of-the-art methods.

Expensive multi-objective optimization problems (EMOPs) are common in real-world scenarios where evaluating objective functions is costly and involves extensive computations or physical experiments. Current Pareto set learning methods for such problems often rely on surrogate models like Gaussian processes to approximate the objective functions. These surrogate models can become fragmented, resulting in numerous small uncertain regions between explored solutions. When using acquisition functions such as the Lower Confidence Bound (LCB), these uncertain regions can turn into pseudo-local optima, complicating the search for globally optimal solutions. To address these challenges, we propose a novel approach called SVH-PSL, which integrates Stein Variational Gradient Descent (SVGD) with Hypernetworks for efficient Pareto set learning. Our method addresses the issues of fragmented surrogate models and pseudo-local optima by collectively moving particles in a manner that smooths out the solution space. The particles interact with each other through a kernel function, which helps maintain diversity and encourages the exploration of underexplored regions. This kernel-based interaction prevents particles from clustering around pseudo-local optima and promotes convergence towards globally optimal solutions. Our approach aims to establish robust relationships between trade-off reference vectors and their corresponding true Pareto solutions, overcoming the limitations of existing methods. Through extensive experiments across both synthetic and real-world MOO benchmarks, we demonstrate that SVH-PSL significantly improves the quality of the learned Pareto set, offering a promising solution for expensive multi-objective optimization problems.

High-quality, large-scale instructions are crucial for aligning large language models (LLMs), however, there is a severe shortage of instruction in the field of natural language understanding (NLU). Previous works on constructing NLU instructions mainly focus on information extraction (IE), neglecting tasks such as machine reading comprehension, question answering, and text classification. Furthermore, the lack of diversity in the data has led to a decreased generalization ability of trained LLMs in other NLU tasks and a noticeable decline in the fundamental model's general capabilities. To address this issue, we propose Hum, a large-scale, high-quality synthetic instruction corpus for NLU tasks, designed to enhance the NLU capabilities of LLMs. Specifically, Hum includes IE (either close IE or open IE), machine reading comprehension, text classification, and instruction generalist tasks, thereby enriching task diversity. Additionally, we introduce a human-LLMs collaborative mechanism to synthesize instructions, which enriches instruction diversity by incorporating guidelines, preference rules, and format variants. We conduct extensive experiments on 5 NLU tasks and 28 general capability evaluation datasets for LLMs. Experimental results show that Hum enhances the NLU capabilities of six LLMs by an average of 3.1%, with no significant decline observed in other general capabilities.

Understanding and predicting the popularity of online User-Generated Content (UGC) is critical for various social and recommendation systems. Existing efforts have focused on extracting predictive features and using pre-trained deep models to learn and fuse multimodal UGC representations. However, the dissemination of social UGCs is not an isolated process in social network; rather, it is influenced by contextual relevant UGCs and various exogenous factors, including social ties, trends, user interests, and platform algorithms. In this work, we propose a retrieval-based framework to enhance the popularity prediction of multimodal UGCs. Our framework extends beyond a simple semantic retrieval, incorporating a meta retrieval strategy that queries a diverse set of relevant UGCs by considering multimodal content semantics, and metadata from user and post. Moreover, to eliminate irrelevant and noisy UGCs in retrieval, we introduce a new measure called Relative Retrieval Contribution to Prediction (RRCP), which selectively refines the retrieved UGCs. We then aggregate the contextual UGC knowledge using vision-language graph neural networks, and fuse them with an RRCP-Attention-based prediction network. Extensive experiments on three large-scale social media datasets demonstrate significant improvements ranging from 26.68% to 48.19% across all metrics compared to strong baselines.

Probability calibration transforms raw output of a classification model into empirically interpretable probability. When the model is purposed to detect rare event and only a small expensive data source has clean labels, it becomes extraordinarily challenging to obtain accurate probability calibration. Utilizing an additional large cheap data source is very helpful, however, such data sources oftentimes suffer from biased labels. To this end, we introduce an approximate expectation-maximization (EM) algorithm to extract useful information from the large data sources. For a family of calibration methods based on the logistic likelihood, we derive closed-form updates and call the resulting iterative algorithm CalEM. We show that CalEM inherits convergence guarantees from the approximate EM algorithm. We test the proposed model in simulation and on the real marketing datasets, where it shows significant performance increases.

Transferable adversarial examples are known to cause threats in practical, black-box attack scenarios. A notable approach to improving transferability is using integrated gradients (IG), originally developed for model interpretability. In this paper, we find that existing IG-based attacks have limited transferability due to their naive adoption of IG in model interpretability. To address this limitation, we focus on the IG integration path and refine it in three aspects: multiplicity, monotonicity, and diversity, supported by theoretical analyses. We propose the Multiple Monotonic Diversified Integrated Gradients (MuMoDIG) attack, which can generate highly transferable adversarial examples on different CNN and ViT models and defenses. Experiments validate that MuMoDIG outperforms the latest IG-based attack by up to 37.3% and other state-of-the-art attacks by 8.4%. In general, our study reveals that migrating established techniques to improve transferability may require non-trivial efforts.

Deep neural networks (DNNs) are susceptible to universal adversarial perturbations (UAPs). These perturbations are meticulously designed to fool the target model universally across all sample classes. Unlike instance-specific adversarial examples (AEs), generating UAPs is more complex because they must be generalized across a wide range of data samples and models. Our research reveals that existing universal attack methods, which optimize UAPs using DNNs with static model parameter snapshots, do not fully leverage the potential of DNNs to generate more effective UAPs. Rather than optimizing UAPs against static DNN models with a fixed training set, we suggest using dynamic model-data pairs to generate UAPs. In particular, we introduce a dynamic maximin optimization strategy, aiming to optimize the UAP across a variety of optimal model-data pairs. We term this approach DM-UAP. DM-UAP utilizes an iterative max-min-min optimization framework that refines the model-data pairs, coupled with a curriculum UAP learning algorithm to examine the combined space of model parameters and data thoroughly. Comprehensive experiments on the ImageNet dataset demonstrate that the proposed DM-UAP markedly enhances both cross-sample universality and cross-model transferability of UAPs. Using only 500 samples for UAP generation, DM-UAP outperforms the state-of-the-art approach with an average increase in fooling ratio of 12.108%.

Recent studies showed that the generalization of neural networks is correlated with the sharpness of the loss landscape and flat minima suggests a better generalization ability than sharp minima. In this paper, we propose a novel method called optimum shifting, which changes the parameters of a neural network from a sharp minimum to a flatter one while maintaining the same training loss value. Our method is based on the observation that when the input and output of a neural network are fixed, the matrix multiplications within the network can be treated as systems of under-determined linear equations, enabling adjustment of parameters in the solution space, which can be simply accomplished by solving a constrained optimization problem. Furthermore, we introduce a practical stochastic optimum shifting technique utilizing the neural collapse theory to reduce computational costs and provide more degrees of freedom for optimum shifting. Extensive experiments with various deep neural network architectures on benchmark datasets demonstrate the effectiveness of our method.

Dealing with the distribution shift is a significant challenge when building offline reinforcement learning (RL) models that can generalize from a static dataset to out-of-distribution (OOD) scenarios. Previous approaches have employed pessimism or conservatism strategies. More recently, data-driven work has taken a distributional perspective, treating offline data as a domain adaptation problem. However, these methods use heuristic techniques to simulate distribution shifts, resulting in a limited diversity of artificially created distribution gaps. In this paper, we propose a novel perspective: offline datasets inherently contain multiple latent distributions, with behavior data from diverse policies potentially following different distributions and data from the same policy across various time phases also exhibiting distribution variance. We introduce the Latent Distribution Representation Learning (LAD) framework, which aims to characterize the multiple latent distributions within offline data and reduce the distribution gaps between any pair of them. LAD consists of a min-max adversarial process: it first identifies the "worst-case" distributions to enlarge the diversity of distribution gaps and then reduces these gaps to learn invariant representations for generalization. We derive a generalization error bound to support LAD theoretically and verify its effectiveness through extensive experiments.

AI-synthesized voice technology has the potential to create realistic human voices for beneficial applications, but it can also be misused for malicious purposes. While existing AI-synthesized voice detection models excel in intra-domain evaluation, they face challenges in generalizing across different domains, potentially becoming obsolete as new voice generators emerge. Current solutions use diverse data and advanced machine learning techniques (e.g., domain-invariant representation, self-supervised learning), but are limited by predefined vocoders and sensitivity to factors like background noise and speaker identity. In this work, we introduce an innovative disentanglement framework aimed at extracting domain-agnostic artifact features related to vocoders. Utilizing these features, we enhance model learning in a flat loss landscape, enabling escape from suboptimal solutions and improving generalization. Extensive experiments on benchmarks show our approach outperforms state-of-the-art methods, achieving up to 5.12% improvement in the equal error rate metric in intra-domain and 7.59% in cross-domain evaluations.

With the popularity of federated learning, federated domain generalization (FedDG) has attracted more and more attentions. Existing works of federated learning indicate that the generalization performance of the global model can be improved when the global model is obtained by aggregating local models according to a suitable weights. However, the existing methods to calculate weights do not fully utilize the data influences on the global model update, which gives us an opportunity to improve the generalization performance of the global model further. In this paper, we propose the method DI (data influences), which utilizes the data influences on the global model update to calculate dynamical weights of local model in each round of training. Specifically, the first component data influences calculator (DIC) of DI calculates the local weights of local model from the influences of each data on the global model update and we introduce the influences function to complete the calculation process. The second component data influences adjuster (DIA) of DI calculates the global weights (which are used in the aggregation process of the global model) from local weights. Extensive experiments indicate that our method improves the generalization performance of models significantly. In particular, our method improves model accuracy on benchmark datasets PACS, OfficeHome, and Office-31 by 1.79%, 1.61%, and 2.39% on average, respectively. Source code is publicly available at github.

Despite their remarkable capabilities, Large Language Models (LLMs) are prone to generate responses that contradict verifiable facts, i.e., unfaithful hallucination content. Existing efforts generally focus on optimizing model parameters or editing semantic representations, which compromise the internal factual knowledge of target LLMs. In addition, hallucinations typically exhibit multifaceted patterns in downstream tasks, limiting the model's holistic performance across tasks. In this paper, we propose a Comparator-driven Decoding-Time (CDT) framework to alleviate the response hallucination. Firstly, we construct hallucinatory and truthful comparators with multi-task fine-tuning samples. In this case, we present an instruction prototype-guided mixture of experts strategy to enhance the ability of the corresponding comparators to capture different hallucination or truthfulness patterns in distinct task instructions. CDT constrains next-token predictions to factuality-robust distributions by contrasting the logit differences between the target LLMs and these comparators. Systematic experiments on multiple downstream tasks show that our framework can significantly improve the model performance and response factuality.

We present theory of synaptic neural balance and we show experimentally that synaptic neural balance can improve deep learning speed, and accuracy, even in data-scarce environments. Given an additive cost function (regularizer) of the synaptic weights, a neuron is said to be in balance if the total cost of its incoming weights is equal to the total cost of its outgoing weights. For large classes of networks, activation functions, and regularizers, neurons can be balanced fully or partially using scaling operations that do not change their functionality. Furthermore, these balancing operations are associated with a strictly convex optimization problem with a single optimum and can be carried out in any order. In our simulations, we systematically observe that: (1) Fully balancing before training results in better performance as compared to several other training approaches; (2) Interleaving partial (layer-wise) balancing and stochastic gradient descent steps during training results in faster learning convergence and better overall accuracy (with L1 balancing converging faster than L2 balancing); and (3) When given limited training data, neural balanced models outperform plain or regularized models; and this is observed in both feedforward and recurrent networks. In short, the evidence supports that neural balancing operations could be added to the arsenal of methods used to regularize and train neural networks. Furthermore, balancing operations are entirely local and can be carried out asynchronously, making them plausible for biological or neuromorphic systems.