Beyond Mean-Variance: AI Takes the Reins of Portfolio Management
![Modern portfolio optimization, traditionally reliant on mean-variance optimization [latex]MVO[/latex], is shown to be outperformed by a reinforcement learning approach [latex]DRL[/latex] during backtesting, suggesting that algorithms mirroring human adaptability-even with inherent imperfections-can navigate market fluctuations more effectively than static, mathematically idealized models.](https://arxiv.org/html/2602.17098v1/figures/backtest.png)
A new study reveals that deep reinforcement learning consistently delivers superior portfolio performance compared to traditional optimization techniques.
Science
Building Better Graph Networks with Adversarial Training
![The AdvSynGNN framework achieves resilient node labeling through a process of iterative refinement, beginning with multi-scale feature synthesis and progressing via contrastive representation alignment-stabilized by a self-supervised loss [latex]\mathcal{L}_{ssl}[/latex]-and adversarial perturbation of the graph structure with heterophily-oriented edge flips, followed by adaptive residual correction utilizing per-node calibration [latex]c_i[/latex] to mitigate noise, and culminating in a heterophily-adaptive graph transformer incorporating learned structural attention [latex]\phi_{ij}[/latex] before a robust diffusion module computes a steady-state prediction [latex]Z^{(\in fty)}[/latex] integrated through prediction fusion and ensemble to yield final labels [latex]Y_{final}[/latex], with jointly optimized modules driving end-to-end training.](https://arxiv.org/html/2602.17071v1/figure/AdvSynGNN.png)
A new approach, AdvSynGNN, boosts the performance and resilience of graph neural networks on complex, real-world datasets.
The Price of Cooperation: When Algorithms Learn to Collude
![This study demonstrates that a meta-game designed for repeated Prisoner’s Dilemma, employing canonical strategies, converges towards cooperative outcomes when analyzed using Q-learning with a discount factor of [latex]\gamma = 0.95[/latex], achieved through pretraining and testing with normalized initial Q-values and a purely exploitative strategy ([latex]\epsilon = 0[/latex]) over [latex]t = 80[/latex] iterations, ensuring convergence of the base games as detailed in Example 3.1.](https://arxiv.org/html/2602.17203v1/x1.png)
New research reveals that artificial intelligence agents can independently develop collusive strategies in repeated interactions, even without explicit programming or communication.
Evolving Strategy: AI Designs New Algorithms for Multi-Agent Games
Researchers are leveraging the power of large language models to automatically create novel learning algorithms, moving beyond parameter tuning to genuine code evolution.
Turning Talk into Tools: Building Smarter AI Assistants
New research shows how to leverage existing conversation data to create AI assistants that handle routine tasks and seamlessly escalate complex issues.
Hunting for Hidden Signals: AI Finds Faint Radio Emissions
A new machine learning approach dramatically improves the detection of faint, diffuse radio emissions within massive astronomical datasets.
Weaving Tales with AI: How Storytelling Principles are Shaping Language Models
This review explores the growing synergy between the art of narrative and the power of artificial intelligence, examining how longstanding theories of storytelling are being applied to advance language model capabilities.
Learning to Predict Material Behavior with Differentiable Density Functional Theory
Researchers have developed a new machine learning framework that optimizes electronic structure calculations for crystalline materials, promising more accurate and efficient simulations.
Seeing is Believing: The Rise of Real-Time Object Detection
Deep learning is rapidly transforming our ability to instantly identify objects in images and videos, powering applications from autonomous vehicles to advanced robotics.
Filling the Gaps in Climate History with AI
![A probabilistic deep learning framework reconstructs Earth system dynamics by first establishing a generative climate prior through unsupervised pre-training on climate simulations and reanalysis data, then guiding spatiotemporally consistent field generation-initiated from noise and solved via reverse-time Stochastic Differential Equations-with simultaneous gradients enforcing fidelity to sparse station data [latex]\mathcal{G}\_{obs}[/latex] and temporal continuity between overlapping time windows [latex]\mathbf{w}\_{i},\mathbf{w}\_{i+1}[/latex], ultimately yielding both high-fidelity reconstructions optimized for fine-scale structures and large-ensemble realizations quantifying uncertainty.](https://arxiv.org/html/2602.16515v1/x23.png)
A new generative AI framework leverages diffusion models to create more accurate and reliable reconstructions of past global climate conditions.