Decoding Finance’s Visual Language
New research reveals the challenges vision-language models face when interpreting complex financial documents, particularly those containing charts and tables.
Science
Seeing Clearly Through the Noise: Robustness in Financial Forecasting
![The adversarial robustness of Lite-CNN models, evaluated through metrics like [latex]\mathrm{MCC}(\epsilon\_{\mathrm{adv}})[/latex] under both FGSM and PGD attacks, demonstrates that late fusion strategies degrade more gracefully than early fusion when subjected to perturbations targeting individual or combined views of the data.](https://arxiv.org/html/2602.11020v1/x5.png)
New research reveals that while combining multiple data sources can improve financial predictions, these systems are surprisingly vulnerable to even minor data manipulation.
Dynamic Networks: Rewiring Graphs for Smarter AI
![A graph neural network framework embraces sparsity as an inherent characteristic, initializing weight matrices with the ε-based Erdős-Rényi method or proceeding without, then optimizing parameters via gradient descent while either maintaining fixed sparsity [latex]\zeta_{f}[/latex] or adaptively rewiring connections with [latex]\zeta_{a}[/latex] throughout training-a process diverging from baseline training and ultimately yielding graph classification.](https://arxiv.org/html/2602.10754v1/x2.png)
A new approach to graph neural networks dynamically adjusts connections to improve performance and efficiency, even under critical conditions.
Predicting Market Moves with Asset Embeddings
A new framework learns to represent assets in a way that anticipates future correlations, improving portfolio construction and risk management.
Predicting Stellar Collisions with the Power of Machine Learning
![The study systematically mapped the outcomes of stellar collisions across a comprehensive range of primary and secondary masses, revealing how the resulting stellar remnants-whether fully merged or stripped of material-varied with collision parameters like pericenter distance [latex]r_{p}[/latex] and velocity at infinity [latex]v_{\in fty}[/latex], ultimately demonstrating the complex interplay of these factors in shaping post-collision stellar evolution.](https://arxiv.org/html/2602.10191v1/x1.png)
New research leverages machine learning models trained on detailed simulations to rapidly and accurately forecast the outcomes of stellar collisions.
Seeing Through the Smoke and Mirrors: Improving Deepfake Detection Reliability
New research explores how quantifying uncertainty can help deepfake detection systems identify when they don’t know, leading to more trustworthy results.
AI’s New Social Life: Inside Moltbook
Researchers have published the first comprehensive analysis of Moltbook, a social network populated entirely by artificial intelligence agents, revealing a surprisingly complex and sometimes troubling digital ecosystem.
Beyond Text: Evaluating Financial AI’s Multilingual, Multimodal Reasoning
A new benchmark is emerging to rigorously assess the ability of advanced AI systems to understand and reason about financial data presented in multiple languages and formats.
Mapping Eurozone Banking Risk
A new network model integrates detailed financial data to reveal hidden connections and vulnerabilities within the euro area’s banking system.
AI Predicts Material Magnetism with Unprecedented Accuracy
![A neural network variational Monte Carlo technique optimizes a representation of a quantum system’s ground state-defined by a first-principles Hamiltonian-through energy minimization, identifying magnetic order via total and staggered magnetization calculations; this approach, when applied to a moiré semiconductor model of WSe2/WS2 with 2828 electrons, reveals an itinerant ferromagnetic state characterized by spin density and visualized through spin polarization-quantified as [latex]\frac{\langle\hat{n}\_{\uparrow}\rangle-\langle\hat{n}\_{\downarrow}\rangle}{\langle\hat{n}\_{\uparrow}\rangle+\langle\hat{n}\_{\downarrow}\rangle}[/latex]-and total density.](https://arxiv.org/html/2602.09093v1/x1.png)
A new machine learning approach leverages first-principles calculations to accurately forecast magnetic properties, paving the way for the discovery of novel materials.