Alignment techniques often inadvertently induce sycophancy in LLMs.While prior studies studied this behaviour in direct-answer settings, the role of Chain-of-Thought (CoT) reasoning remains under-explored: does it serve as a logical constraint that mitigates sycophancy, or a tool for post-hoc rationalization that masks it?We evaluate a range of models across objective and subjective tasks to investigate the issue.Results show that reasoning generally reduces sycophancy in final decisions but also masks sycophancy in some samples, where models construct deceptive justifications through logical inconsistencies, calculation errors, and onesided arguments, etc.Furthermore, LLMs are more prone to sycophancy in subjective tasks and under authority-bias.Our mechanistic analysis on three open-source models reveals that the tendency of sycophancy is dynamic during the reasoning process rather than being predetermined at the input stage 1 .Nostalgebraist.

Unmanned aerial vehicles (UAVs) have been applied to various tasks in the low-altitude economy (LAE) with the advantages of high mobility, low costs, and flexible deployment. However, due to the broadcast nature of wireless channels and the increasing number of UAVs, the security of UAV control information and the spectrum resource utilization face significant challenges and threats. Therefore, in this paper, we investigate the secrecy performance of UAV short-packet control information transmission networks based on rate-splitting multiple access (RSMA) with the presence of multiple eavesdroppers. Moreover, we consider and analyze the impacts of both imperfect channel state information (CSI) and successive interference cancellation (SIC) in a more realistic scenario. Considering both large-scale fading and Nakagami-msmall-scale fading, the closed-form expression of the average effective secrecy sum rate is derived utilizing stochastic geometry and the Gauss-Chebyshev quadrature. Considering that the private stream can be concealed within the high-power common stream, an optimization problem is formulated to maximize the common rate by jointly optimizing the blocklength and power allocation coefficients to enhance security. The block coordinate descent (BCD) algorithm is adopted to solve this problem. Finally, simulation results demonstrate the accuracy of the analysis and the effectiveness of the proposed scheme.

Unmanned aerial vehicles (UAVs) assisted communications have become one of the crucial approaches to enable the reliable and flexible data transmissions, particularly in ultra-reliable and low-latency scenarios, such as remote sensing, emergency response, and military long-range command transmission. In this paper, we investigate the secrecy performance of UAV-assisted short-packet transmission via non-orthogonal multiple access (NOMA), where a UAV serves as an aerial relay to forward mission-critical information from a base station to two remote users in the presence of a ground-based eavesdropper. Both the base station and UAV relay use beamforming for generating the artificial noise to disrupt the eavesdropping and enhance the security, and the UAV operates in half-duplex mode to meet resource constraints and avoid self-interference. The weighted effective secrecy rates of the two users are maximized by jointly optimizing the blocklength, transmission rate, power allocation coefficients, power-sharing factors and UAV position, which is shown to be non-convex and difficult to be solved directly. Accordingly, we decompose the problem into four sub-problems by applying the block coordinate descent (BCD) algorithm to maximize the weighted effective secrecy rate. Then, slack variables are introduced to further solve the sub-problems via successive convex approximation (SCA). Finally, simulation results are presented to demonstrate the effectiveness of the proposed scheme.

Autoregressive Large Language Models (LLMs) demonstrate exceptional performance in language understanding and generation. However, their application in text embedding tasks has been relatively slow, along with the analysis of their semantic representation in probing tasks, due to the constraints of the unidirectional attention mechanism. This paper aims to explore whether such constraints can be overcome by enabling bidirectional attention in LLMs. We tested different variants of the Llama architecture through additional training steps, progressively enabling bidirectional attention and unsupervised/supervised contrastive learning. Our results show that bidirectional attention improves the LLMs’ ability to represent subsequent context but weakens their utilization of preceding context, while contrastive learning training can help to maintain both abilities.

Effectively resolving phonological ambiguities is crucial for robust natural language processing, as these ambiguities are pervasive in tasks ranging from speech-to-text, spelling correction, to offensive language detection.However, current Large Language Models (LLMs) frequently struggle to resolve such ambiguities.To address this challenge, we present a framework to enhances LLMs' phonological capability through a multiple-stage training approach.Our method begins with supervised fine-tuning on well-constructed datasets, including three subtask datasets designed to enhance the model's foundational phonological knowledge, along with a synthetic dataset of step-by-step reasoning chains.Following this, we apply reinforcement learning to incentivize and stabilize its reasoning.Results show that our framework enables the base model to achieve relatively comparable performance to a much larger model.Our ablation studies reveal that subtask datasets and the synthetic dataset can simultaneously impact as complementary modular enhancers to strengthen LLMs' integrated application 1 .

Unmanned aerial vehicle (UAV) assisted wireless communication is essential for the next-generation mobile networks. In coping with the increased dynamics in UAV networks, the design of control information transmission is essential, requiring ultra reliability, low latency, and high security. In this paper, considering both the large-scale path loss and the Nakagami-m small-scale fading, we investigate the secrecy performance of UAV control information transmission in a NOMA ground-air network with an external flying eavesdropper. A spherical secrecy protection zone is set, and the closed-form expressions for average secure BLER and average achievable secrecy throughput are derived. After that, the asymptotic performance in the high SNR regime is analyzed to get more insights. Ultimately, simulation results verify the accuracy of analysis.

Unmanned aerial vehicle (UAV) assisted wireless communication is a key component of the next-generation mobile networks. In coping with the increased dynamics in UAV networks, the transmission of control information is indispensable, requiring not only ultra reliability and low latency, but also high security. In this paper, we investigate the secrecy performance of the control information in a NOMA ground-air short-packet wireless network with an untrusted internal UAV or an external flying eavesdropper, respectively. Both the large-scale path loss and the Nakagami-m small-scale fading are considered. First, the closed-form expressions of the average secure block error rate (BLER) and the average achievable secrecy throughput in each scenario are derived. Then, the asymptotic performance in the high signal-to-noise ratio (SNR) regime is analyzed to get more insights from both scenarios. Specifically, analytical results show that error floors occur with the increase of SNR. Moreover, a one-dimensional search is applied to maximize the average achievable secrecy throughput by optimizing the blocklength. Simulation results are provided to verify the accuracy of analysis and the effectiveness of optimization.

Preschool children have difficulty learning and comprehending abstract concepts, and the cognition of numbers has always been the key to mathematical enlightenment for young children. Our research aims to help preschool children build their cognition of cardinal and ordinal numbers, comprehend simple logical relationships, and master simple digital addition. We developed an Augmented Reality learning tool based on theories related to number cognition and a theoretical framework of software design for preschool children. We also conducted a teaching experiment in a kindergarten, and interviewed the teachers of the kindergarten to learn about their attitudes towards the application of AR in preschool education. Through data analysis, interviews, and discussions, we conclude that (a) AR application can positively influence children’s cognitive digital skills; (b) children have positive attitudes and positive evaluations toward AR application use, but there are some unavoidable problems in children’s attention allocation; (c) proficiency in operating AR tools has a large impact on children’s learning effects.