Cracking the Code: How Attackers Scale Up to Bypass AI Safety
![The study demonstrates that jailbreak success against Llama-3.1-8B-Instruct exhibits diminishing returns with increased attack compute (measured in FLOPs), following a saturating exponential relationship formalized in [latex]Eq. (7)[/latex], as evidenced by the convergence of average red-team scores (ASR) despite escalating computational effort.](https://arxiv.org/html/2603.11149v1/x1.png)
New research reveals predictable patterns in how effectively attackers can circumvent safeguards in large language models, offering a surprising look at the economics of AI security.
Science
Phonon Engineering for Enhanced Electron Binding in 1D Materials
![The study demonstrates that in a doped Harper-Shapira-Shultz (HSSH) model-a 90-site one-dimensional chain with [latex]U=8t[/latex], [latex]\Omega=t[/latex], and doping levels of 6.67% or with two localized defects-the binding energy [latex]\Delta_b[/latex] exhibits a strong dependence on the ratio [latex]\Omega^{\prime}/\Omega[/latex], revealing distinct behaviors between the HSSH and Harper-Haldane (HH) models with parameters [latex]g=0.4[/latex] and [latex]g=1.13[/latex] respectively, and directly correlating with the system’s lattice distortion and phonon dispersion.](https://arxiv.org/html/2603.11373v1/x1.png)
New research reveals how manipulating optical phonons can strengthen electron pairing and induce novel correlated states in one-dimensional doped materials.
Taming the Electron-Phonon Beast
![For an 8-site, 8-electron system, a density matrix renormalization group (DMRG) and functional-renormalization-based embedding method accurately trace the ground state energy across weak and strong coupling regimes, with observed deviations at low [latex]U[/latex] and high [latex]g[/latex] attributable to mean-field overstabilization of charge density wave order-an effect that dissipates within the Mott insulating phase.](https://arxiv.org/html/2603.11463v1/x5.png)
A novel computational approach efficiently simulates the complex interplay between electrons and vibrations in strongly correlated materials.
Wormholes Under the Microscope: Quantum Effects and Stability
![The geometry of a wormhole, described by [latex] \mathbb{R} \times S^{1} [/latex], is visualized as a cylindrical structure of length [latex] L_{0} [/latex] and radius [latex] a [/latex], where the boundaries-circles of radius [latex] a [/latex] at [latex] z = \pm L_{0}/2 [/latex]-represent the points of contact between the flat Minkowski spaces and the wormhole itself.](https://arxiv.org/html/2603.11724v1/x1.png)
New research explores how quantum fluctuations of energy can impact the viability of traversable wormholes, challenging our understanding of these theoretical shortcuts through spacetime.
The Flavor Lattice: A New Route to Quark and Lepton Hierarchies
A novel framework leveraging lattice structures and discrete symmetries offers a compelling explanation for the observed patterns in fundamental particle masses and mixing.
When AI Agents Go Rogue: Securing the Future of Autonomous Systems
As increasingly sophisticated AI agents take on more complex tasks, understanding and mitigating their inherent security risks is paramount.
Twisted Spins: A New Route to Robust Spintronics
Researchers have discovered a way to create and maintain persistent spin textures in a quasi-two-dimensional material, opening doors for energy-efficient spintronic devices.
Negotiating for Better Requirements: An AI-Powered Approach
A new framework uses artificial intelligence and multi-agent negotiation to balance competing priorities in the complex process of defining software and system requirements.
Twisted Layers, Shared Secrets: Simulating High-Temperature Superconductivity
Researchers are exploring moiré patterns in van der Waals materials to create a unified platform for studying the complex physics of both copper- and iron-based superconductors.
Smarter Text Chunks, Better Answers: A New Approach to Knowledge Retrieval
Researchers have developed a system that intelligently breaks down complex texts to improve the accuracy and relevance of information retrieved for question answering.