Large language models (LLMs) used for multiple-choice and pairwise evaluation tasks often exhibit selection bias due to non-semantic factors like option positions and label symbols.Existing inference-time debiasing is costly and may harm reasoning, while pointwise training ignores that the same question should yield consistent answers across permutations.To address this issue, we propose Permutation-Aware Group Relative Policy Optimization (PA-GRPO), which mitigates selection bias by enforcing permutation-consistent semantic reasoning.PA-GRPO constructs a permutation group for each instance by generating multiple candidate permutations, and optimizes the model using two complementary mechanisms: (1) cross-permutation advantage, which computes advantages relative to the mean reward over all permutations of the same instance, and (2) consistency-aware reward, which encourages the model to produce consistent decisions across different permutations.Experimental results demonstrate that PA-GRPO outperforms strong baselines across seven benchmarks, substantially reducing selection bias while maintaining high overall performance.The code is available on GitHub.

Wireless federated learning (FL) facilitates collaborative training of artificial intelligence (AI) models to support ubiquitous intelligent applications at the wireless edge. However, the inherent constraints of limited wireless resources inevitably lead to unreliable communication, which poses a significant challenge to wireless FL. To overcome this challenge, we propose Sign-Prioritized FL (SP-FL), a novel framework that improves wireless FL by prioritizing the transmission of important gradient information through uneven resource allocation. Specifically, recognizing the importance of descent direction in model updating, we transmit gradient signs in individual packets and allow their reuse for gradient descent if the remaining gradient modulus cannot be correctly recovered. To further improve the reliability of transmission of important information, we formulate a hierarchical resource allocation problem based on the importance disparity at both the packet and device levels, optimizing bandwidth allocation across multiple devices and power allocation between sign and modulus packets. To make the problem tractable, the one-step convergence behavior of SP-FL, which characterizes data importance at both levels in an explicit form, is analyzed. We then propose an alternating optimization algorithm to solve this problem using the Newton-Raphson method and successive convex approximation (SCA). Simulation results confirm the superiority of SP-FL, especially in resource-constrained scenarios, demonstrating up to 9.96% higher testing accuracy on the CIFAR-10 dataset compared to existing methods.

Large language models (LLMs) often struggle to accurately read and comprehend extremely long texts.Current methods for improvement typically rely on splitting long contexts into fixed-length chunks.However, fixed truncation risks separating semantically relevant content, leading to ambiguity and compromising accurate understanding.To overcome this limitation, we propose a straightforward approach for dynamically separating and selecting chunks of long context, facilitating a more streamlined input for LLMs.In particular, we compute semantic similarities between adjacent sentences, using lower similarities to adaptively divide long contexts into variable-length chunks.We further train a question-aware classifier to select sensitive chunks that are critical for answering specific questions.Experimental results on both singlehop and multi-hop question-answering benchmarks show that the proposed approach consistently outperforms strong baselines.Notably, it maintains robustness across a wide range of input lengths, handling sequences of up to 256k tokens.Our datasets and

Unreliable communication is a critical bottleneck for the performance of federated learning (FL) in resourceconstrained wireless networks. To address this issue, we propose a priority-aware transmission strategy, where wireless resources are allocated preferentially based on the importance of data. Specifically, recognizing the crucial role of gradient direction in model updating, we transmit the sign and the modulus of local gradients separately, enabling the reuse of sign packets in the event of erroneous modulus transmission. Furthermore, we introduce a hierarchical resource allocation strategy in the proposed framework, prioritizing key gradients via bandwidth allocation across devices and the sign packet via power allocation at each device. Building upon the theoretical one-step convergence analysis, we formulate the resource allocation optimization problem in an explicit form, which facilitates an alternating optimization algorithm respectively applying the Newton method and technique of successive convex approximation (SCA). Numerical results show the superiority of the proposed scheme in both accuracy and convergence rate compared to existing baselines.

Compared to traditional map-based, goal-driven navigation methods , deep reinforcement learning (DRL)-based goal-driven navigation for mobile robots offers the advantage of not relying on prior map information, it enables autonomous decision-making through continuous interaction with the environment. However, DRL-based goal-driven navigation faces significant challenges in terms of low generalization ability and learning inefficiency. In this paper, we propose a DRL approach , variational sparsity Q network (VSQN), which leverages variational inference and transfer learning to achieve efficient goal-driven navigation. The variational inference framework models weight uncertainty within the network, thereby enhancing the agent’s generalization capability. Furthermore, a hierarchical learning network framework is adopted, and transfer learning is employed to incorporate prior knowledge from a pre-trained model into new navigation tasks . This enables the agent to rapidly adapt to novel tasks without the need for fine-tuning after selecting an optimal sub-goal. This improves the agent’s initial performance in previously unseen navigation tasks . The experimental results indicate that the proposed method achieves a success rate (SR) of 76% and a success weighted by inverse path length (SPL) of 0.52 in previously unencountered environments and target locations within the grid environment , and an SR of 81% with an SPL of 0.30 in the AI2-THOR environment. These findings demonstrate that the method substantially enhances the agent’s generalization capability.

Over-the-air computation (AirComp) has emerged as an essential approach for enabling communication-efficient federated learning (FL) over wireless networks. Nonetheless, the inherent analog transmission mechanism in AirComp-based FL (AirFL) intensifies challenges posed by potential Byzantine attacks. In this paper, we propose a novel Byzantine-robust FL paradigm for over-the-air transmissions, referred to as federated learning with secure adaptive clustering (FedSAC). FedSAC aims to protect a portion of the devices from attacks through zero trust architecture (ZTA) based Byzantine identification and adaptive device clustering. By conducting a one-step convergence analysis, we theoretically characterize the convergence behavior with different device clustering mechanisms and uneven aggregation weighting factors for each device. Building upon our analytical results, we formulate a joint optimization problem for the clustering and weighting factors in each communication round. To facilitate the targeted optimization, we propose a dynamic Byzantine identification method using historical reputation based on ZTA. Furthermore, we introduce a sequential clustering method, transforming the joint optimization into a weighting optimization problem without sacrificing the optimality. To optimize the weighting, we capitalize on the penalty convex-concave procedure (P-CCP) to obtain a stationary solution. Numerical results substantiate the superiority of the proposed FedSAC over existing methods in terms of both test accuracy and convergence rate.

Task-oriented integrated sensing, communication, and computation (ISCC) is a key technology for achieving low-latency edge inference and enabling efficient implementation of artificial intelligence (AI) in industrial cyber-physical systems (ICPS). However, the constrained energy supply at edge devices has emerged as a critical bottleneck. In this paper, we propose a novel energy-efficient ISCC framework for AI inference at resource-constrained edge devices, where adjustable split inference, model pruning, and feature quantization are jointly designed to adapt to diverse task requirements. A joint resource allocation design problem for the proposed ISCC framework is formulated to minimize the energy consumption under stringent inference accuracy and latency constraints. To address the challenge of characterizing inference accuracy, we derive an explicit approximation for it by analyzing the impact of sensing, communication, and computation processes on the inference performance. Building upon the analytical results, we propose an iterative algorithm employing alternating optimization to solve the resource allocation problem. In each subproblem, the optimal solutions are available by respectively applying a golden section search method and checking the Karush-Kuhn-Tucker (KKT) conditions, thereby ensuring the convergence to a local optimum of the original problem. Numerical results demonstrate the effectiveness of the proposed ISCC design, showing a significant reduction in energy consumption of up to 40% compared to existing methods, particularly in low-latency scenarios.

Short videos are one of the most popular social media in the world, triggering a proliferation of copycat creations leading to homogenized video content, with visual and textual homogenization being the most prevalent. Unlike near-duplicate video retrieval, which relies on visual appearance similarity, homogenization recognition emphasizes identifying videos with similar semantic units. Short videos exhibit multimodal features, in which there is a many-to-many mapping relationship between visual and text elements, and the two modalities are relatively independent and semantically correlated. Therefore, cross-modal semantic correlation needs to be explored and established to achieve homogenization recognition of short videos. Based on the idea of divide-and-conquer and joint processing, we propose a cross-modal tri-semantic correlation-CLIP (CS 3 C-CLIP) for short video homogenization recognition. First, visual and text features in the shared subspace are extracted using the contrastive language-image pre-training visual-text dual encoder. Then, features at the patch, frame, and video levels are generated using the patch selection module and the temporal encoder, while the word-level and sentence-level features are respectively derived from text features and [EOS] token. After establishing cross-modal tri-semantic correlations by constructing a triple semantic (i.e., video-sentence, frame-sentence, and patch-word) correlation, homogenized short videos are recognized by measuring the aggregated cross-modal similarity between pairs of short videos. Experimental results on three publicly available datasets demonstrate that our CS 3 C-CLIP outperforms state-of-the-art methods, achieving 85.7% R@1 and 94.4% R@5 on self-built BJUT-HCD, 49.4% R@1 and 74.6% R@5 on MSR-VTT, and 49.8% R@1 and 78.1% R@5 on MSVD, respectively.

In this paper, we quantitatively compare these two effective communication schemes, i.e., digital and analog ones, for wireless federated learning (FL) over resource-constrained networks, highlighting their essential differences as well as their respective application scenarios. We first examine both digital and analog transmission methods, together with a unified and fair comparison scheme under practical constraints. A universal convergence analysis under various imperfections is established for FL performance evaluation in wireless networks. These analytical results reveal that the fundamental difference between the two paradigms lies in whether communication and computation are jointly designed or not. The digital schemes decouple the communication design from specific FL tasks, making it difficult to support simultaneous uplink transmission of massive devices with limited bandwidth. In contrast, the analog communication allows over-the-air computation (AirComp), thus achieving efficient spectrum utilization. However, computation-oriented analog transmission reduces power efficiency, and its performance is sensitive to computational errors. Finally, numerical simulations are conducted to verify these theoretical observations.

Reconfigurable intelligent surface (RIS) is regarded as an important enabling technology for the sixth-generation (6G) network. Recently, modulating information in reflection patterns of RIS, referred to as reflection modulation (RM), has been proven in theory to have the potential of achieving higher transmission rate than existing passive beamforming (PBF) schemes of RIS. To fully unlock this potential of RM, we propose a novel superimposed RIS-phase modulation (SRPM) scheme for multiple-input multiple-output (MIMO) systems, where tunable phase offsets are superimposed onto predetermined RIS phases to bear extra information messages. The proposed SRPM establishes a universal framework for RM, which retrieves various existing RM-based schemes as special cases.Moreover, the advantages and applicability of the SRPM in practice is also validated in theory by analytical characterization of its performance in terms of average bit error rate (ABER) and ergodic capacity. To maximize the performance gain, we formulate a general precoding optimization at the base station (BS) for a single-stream case with uncorrelated channels and obtain the optimal SRPM design via the semidefinite relaxation (SDR) technique. Furthermore, to avoid extremely high complexity in maximum likelihood (ML) detection for the SRPM, we propose a sphere decoding (SD)-based layered detection method with near-ML performance and much lower complexity. Numerical results demonstrate the effectiveness of SRPM, precoding optimization, and detection design. It is verified that the proposed SRPM achieves a higher diversity order than that of existing RM-based schemes and outperforms PBF significantly especially when the transmitter is equipped with limited radio-frequency (RF) chains.

To enable wireless federated learning (FL) in communication resource-constrained networks, two communication schemes, i.e., digital and analog ones, are effective solutions. In this paper, we quantitatively compare these two techniques, highlighting their essential differences as well as respectively suitable scenarios. We first examine both digital and analog transmission schemes, together with a unified and fair comparison framework under imbalanced device sampling, strict latency targets, and transmit power constraints. A universal convergence analysis under various imperfections is established for evaluating the performance of FL over wireless networks. These analytical results reveal that the fundamental difference between the digital and analog communications lies in whether communication and computation are jointly designed or not. The digital scheme decouples the communication design from FL computing tasks, making it difficult to support uplink transmission from massive devices with limited bandwidth and hence the performance is mainly communication-limited. In contrast, the analog communication allows over-the-air computation (AirComp) and achieves better spectrum utilization. However, the computation-oriented analog transmission reduces power efficiency, and its performance is sensitive to computation errors from imperfect channel state information (CSI). Furthermore, device sampling for both schemes are optimized and differences in sampling optimization are analyzed. Numerical results verify the theoretical analysis and affirm the superior performance of the sampling optimization.

In this paper, we consider the robust beamforming design in a reconfigurable intelligent surface (RIS)-aided cell-free (CF) system considering the channel state information (CSI) uncertainties of both the direct channels and cascaded channels at the transmitter with capacity-limited backhaul. We jointly optimize the precoding at the access points (APs) and the phase shifts at multiple RISs to maximize the worst-case sum rate of the CF system subject to the constraints of maximum transmit power of APs, unit-modulus phase shifts, limited backhaul capacity, and bounded CSI errors. By applying a series of transformations, the non-smoothness and semi-infinite constraints are tackled in a low-complexity manner that facilitates the design of an alternating optimization (AO)-based iterative algorithm. The proposed algorithm divides the considered problem into two subproblems. For the RIS phase shifts optimization subproblem, we exploit the penalty convex-concave procedure (P-CCP) to obtain a stationary solution and achieve effective initialization. For precoding optimization subproblem, successive convex approximation (SCA) is adopted with a convergence guarantee to a Karush-Kuhn-Tucker (KKT) solution. Numerical results demonstrate the effectiveness of the proposed robust beamforming design, which achieves superior performance with low complexity. Moreover, the importance of RIS phase shift optimization for robustness and the advantages of distributed RISs in the CF system are further highlighted.

As an emerging field of network content production, live video has been in the vacuum zone of cyberspace governance for a long time. Streamer action recognition is conducive to the supervision of live video content. In view of the diversity and imbalance of streamer actions, it is attractive to introduce few-shot learning to realize streamer action recognition. Therefore, a meta-learning paradigm and CosAttn for streamer action recognition method in live video is proposed, including: (1) the training set samples similar to the streamer action to be recognized are pretrained to improve the backbone network; (2) video-level features are extracted by R(2+1)D-18 backbone and global average pooling in the meta-learning paradigm; (3) the streamer action is recognized by calculating cosine similarity after sending the video-level features to CosAttn to generate a streamer action category prototype. Experimental results on several real-world action recognition datasets demonstrate the effectiveness of our method.