New research reveals how safety mechanisms are encoded within large language models and demonstrates a method to pinpoint and manipulate the specific components responsible for preventing harmful outputs.
A new analysis of waveform data reveals a pattern of low-magnitude earthquakes in the Kamchatka Peninsula leading up to a significant seismic event.
![The Fission-GRPO framework operates through iterative refinement-initially optimizing a policy [latex]\pi_{\theta}[/latex] across a query distribution [latex]\mathcal{D}[/latex], then isolating error trajectories via a diagnostic simulator [latex]\mathcal{S}_{\phi}[/latex], and finally employing a multiplicative resampling process-governed by a factor [latex]G^{\prime}[/latex]-to steer the policy toward successful recovery paths, embodying a system designed not to prevent decay, but to adaptively reconfigure itself within it.](https://arxiv.org/html/2601.15625v1/x2.png)
New research demonstrates a method for improving the reliability of AI agents by transforming failed actions into valuable learning opportunities.
A new approach to modeling spiking neural networks using hypergraphs promises to dramatically improve how these networks are deployed on specialized neuromorphic hardware.
This industry case study details a practical and secure approach to deploying AI-powered chatbots for small businesses using distributed, cloud-native technologies.
A new machine vision system, built using a custom convolutional neural network, is showing promise in the early detection of skin lesions.
![As semantic density increases-measured by ρ-neural accuracy rapidly declines, evidenced by a sharp decrease in N50N\_{50}, which validates the Orthogonality Constraint by demonstrating that higher densities lead to increased key overlap and subsequent interference; achieving values below [latex]\rho < 0.3[/latex] proved unattainable with realistic fact structures.](https://arxiv.org/html/2601.15313v1/fig_density_vs_collapse.png)
New research reveals a fundamental constraint on how neural networks store information, explaining why they falter when faced with complex, overlapping memories.
A new framework uses artificial intelligence to predict and mitigate risks in live streaming by analyzing patterns of behavior across multiple sessions.
Researchers have developed a lightweight model to better understand gene expression within the physical context of tissues, paving the way for more accurate biological insights.
![The model distinguishes itself from traditional Federated Graph Neural Networks by establishing a communication structure-indicated by sequential exchanges [latex]❶[/latex], [latex]❷[/latex], and [latex]❸[/latex]-between server and clients, fundamentally altering the flow of information during the learning process.](https://arxiv.org/html/2601.15722v1/x1.png)
A new approach drastically cuts communication costs in federated graph learning, enabling more practical and privacy-preserving analysis of complex network data.