论文检索

Changyuan Tian, Zhicong Lu, Shuang Qian, Nayu Liu, Peiguang Li, Li Jin, Leiyi Hu, Zhizhao Zeng, Sirui Wang, Ke Zeng 等

To improve Multi-step Mathematical Reasoning (MsMR) of Large Language Models (LLMs), it is crucial to obtain scalable supervision from the corpus by automatically critiquing mistakes in the reasoning process of MsMR and rendering a final verdict of the problem-solution. Most existing methods rely on crafting high-quality supervised fine-tuning demonstrations for critiquing capability enhancement and pay little attention to delving into the underlying reason for the poor critiquing performance of LLMs. In this paper, we orthogonally quantify and investigate the potential reason — imbalanced evaluation preference, and conduct a statistical preference analysis. Motivated by the analysis of the reason, a novel perplexity-aware reinforcement learning algorithm is proposed to rectify the evaluation preference, elevating the critiquing capability. Specifically, to probe into LLMs' critiquing characteristics, a One-to-many Problem-Solution (OPS) benchmark is meticulously constructed to quantify the behavior difference of LLMs when evaluating the problem solutions generated by itself and others. Then, to investigate the behavior difference in depth, we conduct a statistical preference analysis oriented on perplexity and find an intriguing phenomenon — "LLMs incline to judge solutions with lower perplexity as correct", which is dubbed as imbalanced evaluation preference. To rectify this preference, we regard perplexity as the baton in the algorithm of Group Relative Policy Optimization, supporting the LLMs to explore trajectories that judge lower perplexity as wrong and higher perplexity as correct. Extensive experimental results on our built OPS and existing available critic benchmarks demonstrate the validity of our method.

Haitong Luo, Weiyao Zhang, Suhang Wang, Wenji Zou, Chungang Lin, Xuying Meng, Yujun Zhang

The proliferation of high-quality text from Large Language Models (LLMs) demands reliable and efficient detection methods. While existing training-free approaches show promise, they often rely on surface-level statistics and overlook fundamental signal properties of the text generation process. In this work, we reframe detection as a signal processing problem, introducing a novel paradigm that analyzes the sequence of token log-probabilities in the frequency domain. By systematically analyzing the signal's spectral properties using the global Discrete Fourier Transform (DFT) and the local Short-Time Fourier Transform (STFT), we find that human-written text consistently exhibits significantly higher spectral energy. This higher energy reflects the larger-amplitude fluctuations inherent in human writing compared to the suppressed dynamics of LLM-generated text. Based on this key insight, we construct SpecDetect, a detector built on a single, robust feature from the global DFT: DFT total energy. We also propose an enhanced version, SpecDetect++, which incorporates a sampling discrepancy mechanism to further boost robustness. Extensive experiments show that our approach outperforms the state-of-the-art model while running in nearly half the time. Our work introduces a new, efficient, and interpretable pathway for LLM-generated text detection, showing that classical signal processing techniques offer a surprisingly powerful solution to this modern challenge.

Man Hu, Yahui Ding, Yatao Yang, Liangyu Chen, Yanhao Jia, Shuai Zhao

As backdoor attacks become more stealthy and robust, they reveal critical weaknesses in current defense strategies: detection methods often rely on coarse-grained feature statistics, and purification methods typically require full retraining or additional clean models. To address these challenges, we propose DUP (Detection-guided Unlearning for Purification), a unified framework that integrates backdoor detection with unlearning-based purification. The detector captures feature-level anomalies by jointly leveraging class-agnostic distances and inter-layer transitions. These deviations are integrated through a weighted scheme to identify poisoned inputs, enabling more fine-grained analysis. Based on the detection results, we purify the model through a parameter-efficient unlearning mechanism that avoids full retraining and does not require any external clean model. Specifically, we innovatively repurpose knowledge distillation to guide the student model toward increasing its output divergence from the teacher on detected poisoned samples, effectively forcing it to unlearn the backdoor behavior. Extensive experiments across diverse attack methods and language model architectures demonstrate that DUP achieves superior defense performance in detection accuracy and purification efficacy.

Sanchit Sinha, Guangzhi Xiong, Zhenghao He, Aidong Zhang

Modern vision-language models (VLMs) deliver impressive predictive accuracy yet offer little insight into 'why' a decision is reached, frequently hallucinating facts, particularly when encountering out-of-distribution data. Neurosymbolic frameworks address this by pairing black-box perception with interpretable symbolic reasoning, but current methods extract their symbols solely from task labels, leaving them weakly grounded in the underlying visual data. In this paper, we introduce a multi-agent system - Concept-RuleNet that reinstates visual grounding while retaining transparent reasoning. Specifically, a multimodal concept generator first mines discriminative visual concepts directly from a representative subset of training images. Next, these visual concepts are utilized to condition symbol discovery, anchoring the generations in real image statistics and mitigating label bias. Subsequently, symbols are composed into executable first-order rules by a large language model reasoner agent - yielding interpretable neurosymbolic rules. Finally, during inference, a vision verifier agent quantifies the degree of presence of each symbol and triggers rule execution in tandem with outputs of black-box neural models, predictions with explicit reasoning pathways. Experiments on five benchmarks, including two challenging medical-imaging tasks and three underrepresented natural-image datasets, show that our system augments state-of-the-art neurosymbolic baselines by an average of 5% while also reducing the occurrence of hallucinated symbols in rules by up to 50%.

Magauiya Zhussip, Dmitriy Shopkhoev, Ammar Ali, Stamatios Lefkimmiatis

Large language models (LLMs) have revolutionized AI applications, yet their high computational and memory demands hinder their widespread deployment. Existing compression techniques focus on intra-block optimizations (e.g., low-rank approximation or attention head pruning), while the repetitive layered structure of transformers implies significant inter-block redundancy - a dimension largely unexplored beyond key-value (KV) caching. Inspired by dictionary learning in convolutional networks, we propose a framework for structured weight sharing across transformer layers. Our approach decomposes attention projection matrices (Q, K, V, O) into shared dictionary atoms, reducing the attention module's parameters by 66.7% (e.g., 226.5M -> 75M in a 700M-parameter model) while achieving on-par performance. Unlike complex methods requiring distillation or architectural changes, MASA (Matrix Atom Sharing in Attention) operates as a drop-in replacement - trained with standard optimizers - and represents each layer's weights as linear combinations of shared matrix atoms. Experiments across scales (100M-700M parameters) show that MASA achieves better benchmark accuracy and perplexity than grouped-query attention (GQA), low-rank baselines and recently proposed Repeat-all-over/Sequential sharing at comparable parameter budgets. Ablation studies confirm robustness to the dictionary size and the efficacy of shared representations in capturing cross-layer statistical regularities. Extending to Vision Transformers (ViT), MASA matches performance metrics on image classification tasks with 66.7% fewer attention parameters. By combining dictionary learning strategies with transformer efficiency, MASA offers a scalable blueprint for parameter-efficient models without sacrificing performance. Finally, we investigate the possibility of employing MASA on large pretrained models to reduce their number of parameters without experiencing any significant drop in their performance.

Xinxin Zhu, Ying He, Haowen Hou, Ruichong Zhang, Nianbo Zeng, Yulin Peng, Jiongfeng Fang, F. Richard Yu

Reinforcement Fine-tuning (RFT) methods such as Group Relative Policy Optimization (GRPO) have demonstrated strong capabilities in aligning Large Language Models with human preferences. However, these approaches often suffer from limited data efficiency, necessitating extensive on-policy rollouts to maintain competitive performance. We propose PSPO (Prompt-Level Prioritization and Experience-Weighted Smoothing for Efficient Policy Optimization), a lightweight yet effective enhancement to GRPO that improves training stability and sample efficiency through two complementary techniques. First, we introduce an experience-weighted reward smoothing mechanism, which uses exponential moving averages to track group-level reward statistics for each prompt. This enables more stable advantage estimation across training steps without storing entire trajectories, allowing the model to capture historical reward trends in a lightweight and memory-efficient manner. Second, we adopt a prompt-level prioritized sampling strategy, which is an online data selection method inspired by prioritized experience replay. It dynamically emphasizes higher-impact prompts based on their relative advantages, thereby improving data efficiency. Experiments on multiple mathematical reasoning benchmarks and models show that PSPO achieves comparable or better accuracy than GRPO, while significantly accelerating convergence, and maintaining low computational and memory overhead.

Zhengyu Zhou, Weiwei Liu

Bandit multiple hypothesis testing has broad applications in biological sciences, clinical testing for drug discovery, and online A/B/n testing. The framework utilizes an adaptive sampling strategy for multiple testing which aims to maximize statistical power while ensuring anytime false discovery rate control. This paper proposes a robust approach for bandit multiple testing, allowing for at most an epsilon fraction of arbitrary distribution corruption, as in Huber’s contamination model. Specifically, we introduce two adaptive sampling strategies designed to minimize the number of samples required to exceed a target true positive rate, while providing anytime control over the false discovery rate. We analyze the sample complexity of our proposed methods and perform numerical simulations to demonstrate their efficiency and robustness. Furthermore, we extend our methods to address scenarios where distributions have infinite variance and situations involving multiple agents collaborating on the same bandit task.

Yuan-Ting Zhong, Ting Huang, Xiaolin Xiao, Yue-Jiao Gong

Many optimization tasks involve streaming data with unknown concept drifts, posing a significant challenge as Streaming Data-Driven Optimization (SDDO). Existing methods, while leveraging surrogate model approximation and historical knowledge transfer, are often under restrictive assumptions such as fixed drift intervals and fully environmental observability, limiting their adaptability to diverse dynamic environments. We propose TRACE, a TRAnsferable Concept-drift Estimator that effectively detects distributional changes in streaming data with varying time scales. TRACE leverages a principled tokenization strategy to extract statistical features from data streams and models drift patterns using attention-based sequence learning, enabling accurate detection on unseen datasets and highlighting the transferability of learned drift patterns. Further, we showcase TRACE's plug-and-play nature by integrating it into a streaming optimizer, facilitating adaptive optimization under unknown drifts. Comprehensive experimental results on diverse benchmarks demonstrate the superior generalization, robustness, and effectiveness of our approach in SDDO scenarios.

Zijian Zhang, Xinyu Chen, Yuanjie Shi, Liyuan Lillian Ma, Zifan Xu, Yan Yan

Ordinal classification has been widely applied in many high-stakes applications, e.g., medical imaging and diagnosis, where reliable uncertainty quantification (UQ) is essential for decision making. Conformal prediction (CP) is a general UQ framework that provides statistically valid guarantees, which is especially useful in practice. However, prior ordinal CP methods mainly focus on heuristic algorithms or restrictively require the underlying model to predict a unimodal distribution over ordinal labels. Consequently, they provide limited insight into coverage–efficiency trade-offs, or a model-agnostic and distribution-free nature favored by CP methods. To this end, we fill this gap by propose an ordinal-CP method that is model-agnostic and provides instance-level optimal prediction intervals. Specifically, we formulate conformal ordinal classification as a minimum-length covering problem at the instance level. To solve this problem, we develop a sliding-window algorithm that is optimal on each calibration data, with only a linear time complexity in K, the # of label candidates. The local optimality per instance further also improves predictive efficiency in expectation. Moreover, we propose a length-regularized variant that shrinks prediction set size while preserving coverage. Experiments on four benchmark datasets from diverse domains are conducted to demonstrate the significantly improved predictive efficiency of the proposed methods over baselines (by 15%↓ on average over four datasets).

Guanyuan Yu, Yijun Chen, Liang Xu, Gang Kou

Explainability plays a critical role in understanding the workings of Graph Neural Networks (GNNs). While recent methods have introduced causal inference into GNN explanation, they predominantly rely on individual-level interventions and lack rigorous statistical causality testing, resulting in unfaithful and unreliable explanations. To address these challenges, we propose CastX that integrates cohort-level causal analysis with statistical causality testing for GNN explanations. Specifically, CastX formulates the discovery of explanatory subgraphs as a dynamic edge pruning task guided by Conditional Average Treatment Effect (CATE) estimation. A reinforcement learning agent is employed to iteratively eliminate spurious edges and identify causally informative substructures. To further enhance reliability, we introduce an i.i.d.-agnostic non-parametric permutation test that assesses the statistical significance of each target edge. Extensive experiments on real-world datasets demonstrate that our CastX outperforms existing methods in yielding explanatory subgraphs that are concise, faithful, reliable, and statistically supported.

Zhong Ye, Yu Hu, Zhenguo Yang

Multi-task test-time adaptation (MT-TTA) aims to adapt pre-trained models to dynamic environments during multi-task inference by leveraging unlabeled test data. This task is particularly challenging as different tasks respond divergently to distribution shifts, and mixed input streams containing both in-distribution (ID) and out-of-distribution (OOD) samples make the models after test-time adaptation prone to catastrophic forgetting of ID knowledge. Although the existing methods like M-TENT extend the classic test entropy minimization (TENT) by minimizing multi-task entropies and employing task-average gradient to adapt a model, it suffers from two key limitations: 1) the average gradient strategy proposed by M-TENT may exacerbate multi-task test-time optimization conflicts, harming individual tasks when gradients are directionally non-consensual; 2) aggressive updates on mixed ID/OOD data cause severe forgetting of ID knowledge. In this paper, we theoretically establish a formal connection between multi-task loss differences and test-time performance under the first-order Taylor analysis, demonstrating that consensual multi-task entropy reductions are likely to increase the performance, while non-consensual ones might decrease the performance. To this end, we propose Consensus-driven Constrained Multi-Task Test-Time Adaptation (CoCo-MT-TTA), consisting of 1) multi-task gradient consensus adaptation, which aligns cross-task gradient directions to seek a consensus gradient; 2) multi-task plasticity-constraint adaptation, which constrains parameter updates using second-moment statistics to preserve ID knowledge. Extensive experiments on benchmark datasets, including CelebA and PlantData, demonstrate that our method achieves an absolute improvement of up to 16.02% in mean ID/OOD F1-score (Mean I&O) under domain shifts over non-adapted models, outperforming the recent baselines.

Wenhan Wu, Zhili He, Huanghuang Liang, Yili Gong, Jiawei Jiang, Chuang Hu, Dazhao Cheng

Data-protection regulations such as the GDPR grant every participant in a federated system a right to be forgotten. Federated unlearning has therefore emerged as a research frontier, aiming to remove a specific party's contribution from the learned model while preserving the utility of the remaining parties. However, most unlearning techniques focus on Horizontal Federated Learning (HFL), where data are partitioned by samples. In contrast, Vertical Federated Learning (VFL) allows organizations that possess complementary feature spaces to train a joint model without sharing raw data. The resulting feature-partitioned architecture renders HFL-oriented unlearning methods ineffective. In this paper, we propose ReMisVFU, a plug-and-play representation-misdirection framework that enables fast, client-level unlearning in splitVFL systems. When a deletion request arrives, the forgetting party collapses its encoder output to a randomly sampled anchor on the unit sphere, severing the statistical link between its features and the global model. To maintain utility for the remaining parties, the server jointly optimizes a retention loss and a forgetting loss, aligning their gradients via orthogonal projection to eliminate destructive interference. Evaluations on public benchmarks show that ReMisVFU suppresses back-door attack success to the natural class-prior level and sacrifices only about 2.5% points of clean accuracy, outperforming state-of-the-art baselines.

Hong Ye Tan, Subhadip Mukherjee, Junqi Tang, Carola-Bibiane Schönlieb

The manifold hypothesis says that natural high-dimensional data lie on or around a low-dimensional manifold. The recent success of statistical and learning-based methods in very high dimensions empirically supports this hypothesis, suggesting that typical worst-case analysis does not provide practical guarantees. A natural step for analysis is thus to assume the manifold hypothesis and derive bounds that are independent of any ambient dimensions that the data may be embedded in. Theoretical implications in this direction have recently been explored in terms of generalization of ReLU networks and convergence of Langevin methods. In this work, we consider optimal uniform approximations with functions of finite statistical complexity. While upper bounds on uniform approximation exist in the literature using ReLU neural networks, we consider the opposite: lower bounds to quantify the fundamental difficulty of approximation on manifolds. In particular, we demonstrate that the statistical complexity required to approximate a class of bounded Sobolev functions on a compact manifold is bounded from below, and moreover that this bound is dependent only on the intrinsic properties of the manifold, such as curvature, volume, and injectivity radius.

Qi Sun, Zhen Cao, Kaige Geng, Ziyi Zhang, Biao Hou

Spiking Neural Networks (SNNs) offer promising energy efficiency and temporal sparsity for edge intelligence, but their training remains difficult due to gradient mismatch, membrane potential drift, and discretization errors. In this paper, we propose a membrane potential-guided surrogate optimization(MPO) framework that dynamically aligns the surrogate function with the membrane potential distribution to enhance the gradient propagation. Specifically, we introduce a KL-divergence-based regularization to stabilize membrane potential dynamics, and an adaptive width constraint to synchronize the surrogate gradient range with neural activity statistics. Additionally, we design a spike discretization error metric and a correction strategy to mitigate temporal discretization effects. Experiments on CIFAR-10, CIFAR-100, and ImageNet show our method achieves 94.76%, 74.20%, and 65.70% top-1 accuracy respectively, while improving gradient stability and energy efficiency. This work provides a principled optimization scheme for robust and scalable SNN training in practical neuromorphic systems.

Vipul Kumar Singh, Jyotismita Barman, Sandeep Kumar, Tapan K. Gandhi, Jayadeva

Graph Neural Networks (GNNs) are expressive architectures for learning from complex graph-structured data. However, their practical use is often limited by the high computational cost of neighborhood aggregation. Recent efforts have focused on knowledge distillation from GNNs to inference-efficient Multi-Layer Perceptrons (MLPs). However, most existing works treat this distillation as an embedding alignment problem, overlooking the need to replicate the topology-aware smoothing behavior that arises from message passing in GNNs. Moreover, existing methods are primarily performance driven, ignoring critical real-world requirements such as fairness. In this work, we make two key observations: (1) state-of-the-art distillation methods fail to capture the heterogeneous smoothness patterns of GNNs, limiting structural awareness in MLPs, and (2) they introduce significant individual and group fairness violations. We introduce FAITH, the first fair and structurally aware GNN-to-MLP distillation framework with graph-free inference. To improve structural awareness in MLPs, we propose a neighborhood-guided energy alignment objective that transfers not only node-level energy, but also the distribution of energies across local neighborhoods. To improve individual fairness, FAITH introduces a novel ℓ2,1-norm objective that preserves structured similarity in the learned representations. Additionally, we incorporate a counterfactual invariance objective that explicitly encourages the model to learn representations that are statistically independent of the sensitive attribute. We provide a comprehensive theoretical analysis of FAITH, interpreting it through a novel instantiation of the Information Bottleneck principle. Extensive experiments on 11 benchmark datasets show that FAITH achieves stronger structural awareness and delivers a better trade-off between utility and fairness than existing methods.

YongKyung Oh, Dong-Young Lim, Sungil Kim

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.

Pengqian Lu, Jie Lu, Anjin Liu, En Yu, Guangquan Zhang

Existing drift detection methods focus on designing sensitive test statistics. They treat the detection threshold as a fixed hyperparameter, set once to balance false alarms and late detections, and applied uniformly across all datasets and over time. However, maintaining model performance is the key objective from the perspective of machine learning, and we observe that model performance is highly sensitive to this threshold. This observation inspires us to investigate whether a dynamic threshold could be provably better. In this paper, we prove that a threshold that adapts over time can outperform any single fixed threshold. The main idea of the proof is that a dynamic strategy, constructed by combining the best threshold from each individual data segment, is guaranteed to outperform any single threshold that apply to all segments. Based on the theorem, we propose a Dynamic Threshold Determination algorithm. It enhances existing drift detection frameworks with a novel comparison phase to inform how the threshold should be adjusted. Extensive experiments on a wide range of synthetic and real-world datasets, including both image and tabular data, validate that our approach substantially enhances the performance of state-of-the-art drift detectors.

Wenyu Liu, Tianqiang Huang, Pengfei Zhang, Zong Ke, Minghui Min, Puning Zhao

Adversarial attacks pose a major challenge to distributed learning systems, prompting the development of numerous robust learning methods. However, most existing approaches suffer from the curse of dimensionality, i.e. the error increases with the number of model parameters. In this paper, we make a progress towards high dimensional problems, under arbitrary number of Byzantine attackers. The cornerstone of our design is a direct high dimensional semi-verified mean estimation method. The idea is to identify a subspace with large variance. The components of the mean value perpendicular to this subspace are estimated using corrupted gradient vectors uploaded from worker machines, while the components within this subspace are estimated using auxiliary dataset. As a result, a combination of large corrupted dataset and small clean dataset yields significantly better performance than using them separately. We then apply this method as the aggregator for distributed learning problems. The theoretical analysis shows that compared with existing solutions, our method gets rid of sqrt{d} dependence on the dimensionality, and achieves minimax optimal statistical rates. Numerical results validate our theory as well as the effectiveness of the proposed method.

Jaesung Lim, Seunghwan An, Jong-June Jeon

Missing data presents a widespread challenge in real-world data collection. In this paper, our goal is to impute missing entries while accurately reflecting the uncertainty associated with them. We introduce U-VAE, a method that employs a non-parametric distributional learning strategy to parameterize the likelihood of missing values. To address the infeasibility of directly estimating the underlying conditional distributions due to data incompleteness, we incorporate stochastic re-masking and un-masking techniques during training. Specifically, we replace the conventional reconstruction loss with the continuous ranked probability score (CRPS), a strictly proper scoring rule, and theoretically demonstrate that the discrepancy between the underlying conditional distribution and our imputer is upper-bounded. We evaluate the performance of U-VAE on 11 real-world datasets, showing its effectiveness in both single and multiple imputations, while also enhancing post-imputation performance and supporting valid statistical inference.

Gabriele La Malfa, Jie M. Zhang, Michael Luck, Elizabeth Black

Fairness in multi-agent systems (MAS) focuses on equitable reward distribution among agents in scenarios involving sensitive attributes such as race, gender, or socioeconomic status. This paper introduces fairness in Proximal Policy Optimization (PPO) with a penalty term derived from a fairness definition such as demographic parity, counterfactual fairness, or conditional statistical parity. The proposed method, which we call Fair-PPO, balances reward maximisation with fairness by integrating two penalty components: a retrospective component that minimises disparities in past outcomes and a prospective component that ensures fairness in future decision-making. We evaluate our approach in two games: the Allelopathic Harvest, a cooperative and competitive MAS focused on resource collection, where some agents possess a sensitive attribute, and HospitalSim, a hospital simulation, in which agents coordinate the operations of hospital patients with different mobility and priority needs. Experiments show that Fair-PPO achieves fairer policies than PPO across the fairness metrics and, through the retrospective and prospective penalty components, reveals a wide spectrum of strategies to improve fairness; at the same time, its performance pairs with that of state-of-the-art fair reinforcement-learning algorithms. Fairness comes at the cost of reduced efficiency, but does not compromise equality among the overall population (Gini index). These findings underscore the potential of Fair-PPO to address fairness challenges in MAS.