Existing preference optimization objectives for language model alignment require additional hyperparameters that must be extensively tuned to achieve optimal performance, increasing both the complexity and time required fine-tuning large models. In this paper, we propose a simple yet effective hyperparameter-free algorithm alignment.We observe promising performance can achieved simply by optimizing inverse perplexity, which is calculated as of exponentiated average log-likelihood chosen...

The security of global navigation satellite system (GNSS) has attracted a lot attention recently. spoofing detection method using multi-antenna array is one the most efficient methods due to its unique geometry space. However, it either based on assumption that all signals come from same direction or requires additional inertial measurement unit (IMU) attitude solution obtain information. In this paper, we propose new GNSS only two off-the-shelf antennas. This can detect single signal...

The generalization of neural networks is a central challenge in machine learning, especially concerning the performance under distributions that differ from training ones. Current methods, mainly based on data-driven paradigm such as data augmentation, adversarial training, and noise injection, may encounter limited due to model non-smoothness. In this paper, we propose investigate Partial Differential Equation (PDE) perspective, aiming enhance it directly through underlying function...

Consumer-generated messaging through narrowcasting and broadcasting platforms has become more accessible influential. Previous studies have examined why consumers generate messages on different social media from the sender's perspective. This study presents how affective cognitive message framing influence persuasiveness receiver's The findings of Study 1 confirm that is persuasive a platform while platform. mechanism underlies this effect identified in 2, suggesting interaction between...

Real-time system status detection must be accurate and reliable due to the close coupling of Cyber-Physical Systems (CPS) components. In order improve effectiveness CPS anomaly method, this paper proposes a real-time method based on least squares algorithm conditional entropy. To address issue overfitting insufficient generalization squares, penalty term is optimized by two different functions. Meanwhile, K-S test introduced in time window entropy model tackle problem setting threshold. The...

Interval-valued linguistic variables are efficient tools to express the decision makers’ uncertain qualitative judgments. Considering application of interval-valued variables, this paper proposes an interval distance measure, which is then used define measures by combining 2-tuple representation model. To reflect interactions between elements in a set, three correlative on intervalvalued proposed. Meanwhile, several models designed obtain optimal weighting vector constructed. After that,...

This paper introduces a novel generalized self-imitation learning ($\textbf{GSIL}$) framework, which effectively and efficiently aligns large language models with offline demonstration data. We develop $\textbf{GSIL}$ by deriving surrogate objective of imitation density ratio estimates, facilitating the use self-generated data optimizing simple classification losses. eliminates need for complex adversarial training in standard learning, achieving lightweight efficient fine-tuning models. In...

We study the problem of aligning large language models (LLMs) with human preference data. Contrastive optimization has shown promising results in LLMs available data by optimizing implicit reward associated policy. However, contrastive objective focuses mainly on relative values rewards two responses while ignoring their actual values, resulting suboptimal alignment preferences. To address this limitation, we propose calibrated direct (Cal-DPO), a simple yet effective algorithm. show that...

Test-Time Adaptation (TTA) has emerged as a promising paradigm for enhancing the generalizability of models. However, existing mainstream TTA methods, predominantly operating at batch level, often exhibit suboptimal performance in complex real-world scenarios, particularly when confronting outliers or mixed distributions. This phenomenon stems from pronounced over-reliance on statistical patterns over distinct characteristics individual instances, resulting divergence between distribution...

Graph contrastive learning (GCL) emerges as the main-stream approach in graph representation learning, which leverages principle of maximizing mutual information (InfoMax) to learn node representations applied downstream tasks. To explore better generalization from GCL tasks, previous methods heuristically define data augmentation or pretext However, ability and its theoretical are still less reported. In this paper, we first propose a metric called GCL-GE for ability. Considering...

Graph Contrastive Learning (GCL) is increasingly employed in graph representation learning, with the primary aim of developing node/graph representations from a predefined pretext task that can generalize to various downstream tasks. Meanwhile, transition specific diverse and unpredictable tasks poses significant challenge for GCL's generalization ability. Most existing GCL approaches maximize mutual information between two views derived original graphs via data augmentation, either randomly...

The generalization of neural networks is a central challenge in machine learning, especially concerning the performance under distributions that differ from training ones. Current methods, mainly based on data-driven paradigm such as data augmentation, adversarial training, and noise injection, may encounter limited due to model non-smoothness. In this paper, we propose investigate Partial Differential Equation (PDE) perspective, aiming enhance it directly through underlying function...

Face-to-face interactions are central to many individual choices and decision making issues, such as customer service, sales, promotions, negotiations. While the face effect, i.e., face-to-face more effective in inducing compliance than other forms of interactions, has been noted literature, its mechanism rarely explored. This research helps fill theoretical void provides new insights into effect with two lab experiments one field experiment. Study 1, a experiment conducted beauty salon, 2,...

Test-time adaptation (TTA) aims to improve model generalizability when test data diverges from training distribution, offering the distinct advantage of not requiring access and processes, especially valuable in context large pre-trained models. However, current TTA methods fail address fundamental issue: covariate shift, i.e., decreased can be attributed model's reliance on marginal distribution data, which may impair calibration introduce confirmation bias. To this, we propose a novel...

Improving the scalability of GNNs is critical for large graphs. Existing methods leverage three sampling paradigms including node-wise, layer-wise and subgraph sampling, then design unbiased estimator scalability. However, high variance still severely hinders GNNs' performance. On account that previous studies either lacks analysis or only focus on a particular paradigm, we firstly propose an unified node framework analyze core challenge "circular dependency" deriving minimum sampler, i. e.,...

Graph contrastive learning (GCL) emerges as the most representative approach for graph representation learning, which leverages principle of maximizing mutual information (InfoMax) to learn node representations applied in downstream tasks. To explore better generalization from GCL tasks, previous methods heuristically define data augmentation or pretext However, ability and its theoretical are still less reported. In this paper, we first propose a metric named GCL-GE ability. Considering...