Can AI Agents Be Hacked Over Time?
New research reveals that even sophisticated AI agents are vulnerable to subtle, long-term attacks that exploit extended interactions.
Science
Boosting Wireless Security with Intelligent Surfaces
This review explores how reconfigurable intelligent surfaces can be strategically deployed to maximize data security in modern wireless communication networks.
Verifying AI: A Faster Path to Trustworthy Machine Learning
New research introduces Jolt Atlas, a system designed to dramatically accelerate the process of verifying machine learning models using cryptographic techniques.
Hidden Signals: Injecting Persistent Backdoors into AI Vision-Language Models
Researchers have demonstrated a new method for embedding subtle, yet enduring, backdoors into artificial intelligence systems that process both images and text.
Lost in the Code: Why Long Context Isn’t Enough
New research reveals that even as large language models handle increasingly extensive codebases, their ability to reason accurately and avoid distractions remains surprisingly fragile.
Beyond the Codebook: A New Approach to Discrete Representation Learning
![During image reconstruction with [latex]K=1024[/latex], codebook utilization patterns diverge significantly between methods: while VQ-VAE and FSQ initially demonstrate increasing utilization followed by a notable decline, VP-VAE and FSP consistently maintain high and stable codebook engagement throughout the training process.](https://arxiv.org/html/2602.17133v1/x1.png)
A new method decouples learning representations from codebook optimization, addressing common instability issues in vector quantization techniques.
Spin Chain Surprises: When Order Isn’t What It Seems
![The study of coupled Ising chains-specifically, [latex]N=2[/latex] chains with [latex]K=0[/latex] and [latex]g=0.5[/latex]-reveals that two-point correlation functions exhibit critical behavior near phase transitions, evidenced by power-law decay-quantified as [latex]C|i-j|^{-2\Delta}[/latex] for spin correlations-and providing a means to characterize the system's susceptibility to external influence, with reference correlations decaying as [latex]C|i-j|^{-2}[/latex].](https://arxiv.org/html/2602.17029v1/x2.png)
New research reveals unexpected behavior in coupled spin chains, challenging established theories about the transitions between different phases of matter.
Beyond Collisions: Scaling Recommendation with Dynamic Embeddings
A new approach to managing embedding tables minimizes conflicts and ensures models stay up-to-date in large-scale recommendation systems.
Hidden States Complicate Hadron Calculations
![Lattice determinations of the pion-nucleon sigma term [latex]\sigma_{\pi N}[/latex], utilizing both [latex]N_f = 2+1[/latex] and [latex]N_f = 2+1+1[/latex] flavors, converge with phenomenological analyses derived from nucleon-pion scattering-as evidenced by agreement with the FLAG 2024 average-while Mainz lattice group data, obtained from a single ensemble at [latex]m_{\pi} = 200~\rm MeV[/latex] and varying source-sink separations, refine the estimation of [latex]\sigma_{\pi N}[/latex] through two-state fits of summed correlators, further corroborated by ratio analyses utilizing isoscalar scalar current insertions at zero momentum transfer.](https://arxiv.org/html/2602.17195v1/x6.png)
New research reveals that unexpected multi-hadron states induced by calculation methods can significantly skew results in lattice QCD calculations of hadron properties.
Bridging Static and Dynamic Quantum Logic
A new categorical equivalence connects complete orthomodular lattices with a novel class of dynamic algebras, offering fresh insights into the foundations of quantum mechanics.