Evolving Defenses: Neuroevolution Tackles Adversarial Attacks
![The progression of optimal solutions across successive generations reveals a clear inflection point, demarcated by the [latex]\tau_{auth}[/latex] threshold, signifying a notable shift in the evolutionary trajectory.](https://arxiv.org/html/2603.25517v1/x3.png)
A new approach uses genetic algorithms to directly design convolutional neural networks that are more resilient to carefully crafted inputs designed to fool them.
Science
Beyond the Hype: What Users Really Want From AI Chatbots
New research reveals that user satisfaction with leading AI assistants isn’t driven by raw technical power, but by factors like usability and platform integration.
The Rise of the Machine Author
A new analysis reveals how large language models are subtly reshaping academic writing and challenging traditional methods for identifying original research.
Beyond Limits: Encoding the Future with Neural Compression
![A hybrid compression approach-bypassing the neural engine for mixed-modality files-demonstrates that state-of-the-art data density need not come at the expense of speed, effectively mitigating the extreme latency inherent in pure large language model compression methods [latex] (LLMZip) [/latex].](https://arxiv.org/html/2603.25526v1/pareto_efficiency.png)
A new approach to long-term data storage combines the power of large language models with deterministic compression techniques to dramatically improve archival efficiency.
Mapping Neural Network Behavior with Probabilistic Analysis
A new approach uses probabilistic abstract interpretation to track how input data flows through neural networks, offering insights beyond traditional analysis methods.
Smarter Traffic: How AI Drivers Can Ease Congestion
New research explores how reinforcement learning-based automated vehicles can optimize traffic flow and improve fuel efficiency when mixed with human drivers.
Sharper Images, Stronger Defenses: Boosting AI Resilience
A new technique leverages simple image sharpening to significantly improve the robustness of deep learning models against adversarial attacks.
Hidden Signals: Injecting Backdoors into Wireless Communication with Deep Learning
Researchers demonstrate a novel attack that subtly manipulates radio frequency signals to compromise deep learning-based modulation classification systems.
Weather Prediction Gets a Boost from Self-Supervised Learning
A new contrastive learning framework, SPARTA, significantly improves the accuracy and efficiency of weather forecasting and data analysis.
Seeing is No Longer Believing: Tackling Misinformation in Short-Form Video
A new benchmark and AI framework aim to improve the detection and explanation of false claims circulating in the increasingly popular world of micro-videos.