Smarter Seismic Insights: Bridging Expert Knowledge and Machine Learning
![The distributions of extracted score features, conditioned on class labels within simulated environments, demonstrate how a feature summary - constructed from [latex]\ell\ell[/latex]-based scores - characterizes variations in these settings.](https://arxiv.org/html/2604.14809v1/x4.png)
A new framework leverages domain expertise to improve the accuracy and interpretability of seismic event classification, even with incomplete data.
Science
When Logic Breaks Down: Understanding AI Reasoning Errors
New research sheds light on the specific moments large language models stumble during complex reasoning tasks, revealing patterns of failure before they become critical.
false Alarms in Early Warning Systems: A Critical Look at Trend Tests
![The study demonstrates that a modified Mann-Kendall test, applied to time series of lag-1 autocorrelation derived from simulations of a fold normal form ([latex] r = -1 [/latex]) with multiplicative noise, reliably detects trends at the nominal 5% significance level across time series of length [latex] N = 100 [/latex] using rolling windows of relative size α.](https://arxiv.org/html/2604.15230v1/images_annexes/FigB1_Mult_sigma.png)
New research reveals that commonly used statistical methods for predicting abrupt shifts in complex systems are often unreliable due to hidden biases.
Sound Off: Hijacking AI with Hidden Audio Commands
New research reveals that Large Audio-Language Models can be subtly manipulated by imperceptible audio prompts, raising significant security concerns.
Catching Cosmic Explosions: A Real-Time Supernova Hunter
Researchers have developed a new machine learning system to rapidly identify rare and powerful superluminous supernovae from the flood of data generated by modern astronomical surveys.
Life’s Algorithm: The Unexpected Complexity of Fungal Growth
New research reveals that predicting the patterns of even simple fungal networks can be as computationally challenging as solving complex logic puzzles.
Forecasting Power Grid Behavior with AI
A new approach uses machine learning to predict how power systems will respond to changing conditions, offering a faster alternative to traditional simulations.
Power Play: Can Foundation Models Predict Electricity Prices?
A new study benchmarks the performance of advanced time series models against traditional deep learning methods for forecasting day-ahead electricity prices in key European markets.
Seeing Through the Haze: Machine Learning Restores Clarity to Light Signals
New research leverages the power of artificial intelligence to accurately classify and reconstruct structured light beams distorted by atmospheric turbulence.
Orchestrating Giants: Efficient Serving of Massive AI Models
![The system explores opportunities to refine block placement through cache reservation, parameterized by [latex]\mathcal{J}=\{j\_{1},\ldots,j\_{5}}\[/latex], [latex]L=3[/latex], [latex]s\_{m}=1[/latex], [latex]s\_{c}=0.1[/latex], and modulated by block-specific parameters [latex]M\_{j}=3[/latex] if [latex]j=j\_{2}[/latex] and 2 otherwise, alongside timing constraints [latex]\tau^{c}\_{j}=2[/latex] for [latex]j=j\_{2}[/latex] and 1 otherwise, with permissible latency [latex]\tau^{p}\_{j\_{l}}=l\epsilon[/latex] for [latex]0<ϵ≪1[/latex], revealing how algorithmic construction-illustrated for [latex]c=1[/latex]-can be evaluated against the totality of possible chain configurations arising from a given block placement.](https://arxiv.org/html/2604.14993v1/x2.png)
A new approach optimizes resource allocation and load balancing to dramatically reduce response times when deploying large language models in complex, multi-step serving pipelines.