![The PowerFlow framework dynamically adjusts the distribution of logical reasoning-sharpening it with [latex] \alpha > 1 [/latex] to enhance performance or flattening it with [latex] \alpha < 1 [/latex] to encourage creative exploration-resulting in a Pareto improvement over existing approaches to directional capability elicitation.](https://arxiv.org/html/2603.18363v1/x1.png)
A new approach unlocks directional control over language model capabilities, boosting reasoning skills or restoring creative flair without relying on traditional reward-based reinforcement learning.
New research demonstrates a method for large language models to refine their own generation process during inference, leading to higher quality outputs without requiring retraining.
![A novel Meta-Bayesian Federated Learning with Federated Fine-tuning (Meta-BayFLFL) algorithm enables personalized learning across [latex] K[/latex] clients connected to a global server, where each client optimizes a local model-built using Binarized Neural Networks-by adaptively selecting from a range of temporary learning rates, and the server aggregates these refined local models to distribute an updated global model, facilitating efficient and customized learning at scale.](https://arxiv.org/html/2603.18083v1/x2.png)
A novel framework combines Bayesian methods and meta-learning to improve the performance of federated learning systems dealing with messy, real-world data.
A new analysis reveals that so-called ‘spectral’ graph neural networks don’t actually leverage spectral properties for improved performance.
![Transformer-based in-context learning demonstrates robustness across noise distributions-including non-Gaussian and heavy-tailed scenarios like Bernoulli, exponential, Gamma, and Poisson-where the [latex]\ell_{1}[/latex] objective aligns with maximum-likelihood estimation, and even extends to distributions like the Student-t distribution (with [latex]\nu = 2[/latex]) that fall outside traditional finite-variance statistical frameworks, offering a performance comparable to, or exceeding, classical estimators such as least squares, Ridge regression, and [latex]\ell_{1}[/latex] solvers (LP and ADMM).](https://arxiv.org/html/2603.18564v1/figs/500k_steps_student_noise_df_2_noise_std_3_with_data_gamma_and_w_gaussian.png)
New research reveals that Transformer models demonstrate surprising resilience when faced with unexpected changes in data distribution during linear regression, often exceeding the performance of traditional statistical methods.
New research reveals the surprisingly limited, yet focused, spread of political propaganda among AI-driven social media accounts.
This review explores the critical role of negative sampling techniques in building effective search systems, particularly as the field embraces dense retrieval and large language models.
A new analysis reveals that simply adding synthetic data to a dataset doesn’t necessarily improve the reliability of statistical inferences.
Researchers have developed a novel hybrid approach leveraging neural networks to accelerate the computation of energy-minimizing states in the Ginzburg-Landau model, crucial for understanding superconductivity.
A new approach combines the power of deep learning with intuitionistic fuzzy logic to improve the precision of brain MRI image segmentation.