Read today's curated science papers translated into plain English, clear visual diagrams, and working Python code. Explore our daily picks for free, or upgrade to analyze your own PDFs and build a personal library.
Today's picks
Liangkai Liu, Bimal Thapa Magar
arXiv·Nov 1, 2026
Robot safety isn't just about one thing; it's about how camera changes and thinking delays team up to make robots act much more dangerously than either problem alone.
Widely discussed in physical AI safety circles for demonstrating how inference latency and environmental drift compound to create catastrophic risks in robot foundation models, accepted at CoRL 2026 workshops.
T. Daniel Crawford, Aparna Krishnan, Brendan M. Shumberger
arXiv·Nov 10, 2026
For the first time, scientists used coupled-cluster theory to simulate how chiral molecules interact with polarized light, finding that adding special 'core-valence' data to the calculations makes the results much more reliable.
T. Daniel Crawford's group introduces a high-precision coupled-cluster implementation for Vibrational Circular Dichroism (VCD) spectra, providing a new benchmark for simulating chiroptical properties in complex molecular systems.
Fang Wu, Hongbin Lin
arXiv·Nov 2, 2026
Designing new proteins is like finding a needle in a haystack; this paper's ProteoCraft method uses smart hints from AI models to quickly suggest good protein changes, then fully checks only the best ideas, making the search much faster and more successful.
Introduces a tree-based search framework for de novo protein binder design, achieving an 82.9% experimental hit rate for PD-L1 miniproteins and outperforming state-of-the-art diffusion models.
Emanuel Dallas, Mario Morford-Oberst, Eric Weine
arXiv·Oct 28, 2026
Scientists used a quantum computer to show that even when quantum information is protected, too much noise can suddenly make it disappear, like a switch flipping, finding this critical noise point by carefully measuring how 'good' the information was and how 'tangled' the quantum bits became, using clever tricks to ignore the computer's own errors.
First experimental verification of a noise-induced phase transition in quantum circuits, using IBM's Heron processor to show how information protection fails beyond a critical threshold.
Eli Goldin, Alex Lombardi, Tomoyuki Morimae
arXiv·Nov 3, 2026
Quantum advantages like proofs of quantumness and oblivious state preparation don't always mean classical computers are slow, as they can exist even when fundamental computing limits like P=NP collapse.
Proves that proofs of quantumness and oblivious state preparation can coexist with P=NP, showing that quantum advantage does not necessarily imply classical hardness or enable certain cryptographic primitives.
Siyuan Hu, Abulikemu Abuduweili, … Vladlen Koltun
arXiv·Nov 4, 2026
A robot system called HALO learns to do complex tasks like sitting on chairs or pushing boxes by practicing in thousands of fake homes inside a computer, then successfully uses these skills in real homes it has never seen before.
Apple's humanoid autonomy system (HALO) demonstrates zero-shot transfer of complex skills like sitting and crawling from synthetic data to real-world environments using a Unitree G1 robot.
Sybren Drijfhout, Alejandra Sanchez-Franks
arXiv·Nov 9, 2026
New climate models show a big ocean current (AMOC) might suddenly stop, but scientists worry these models might be too sensitive, making the collapse seem more likely than it really is, so they need to check why.
This paper challenges the prevailing view that climate models are too stable, suggesting instead that some CMIP6 models may be predisposed to a spurious AMOC collapse due to high sensitivity.
Ethan Saïman, Feiyu Li, … Victor Balédent
arXiv·Oct 31, 2026
By shining special X-rays on a nickelate crystal, scientists could directly see how electrons are spread out in its key orbitals, finding they are almost equally shared, which helps explain how it becomes superconducting.
Provides the first direct experimental visualization of hole orbitals in bilayer nickelates, a crucial step in understanding the electronic mechanism behind high-temperature superconductivity in these materials.
Oishi Banerjee, Eujin Lee, … Pranav Rajpurkar
arXiv·Nov 4, 2026
Using fake patient photos alongside real X-rays, researchers found that AI models often imagine diseases in the photos to match the X-ray, showing they struggle with conflicting medical evidence.
Introduces an adversarial benchmark using synthetic images to expose how multimodal medical AI models hallucinate when radiographs bias the interpretation of clinical photographs, revealing critical reasoning failures.
Elizabeth G. Campolongo, Eric R. Sokol
arXiv·Nov 2, 2026
This paper introduces AI-FAIR criteria and a Question–Data–Model triangle to make scientific data truly ready for AI, ensuring that new insights from AI models flow back to improve the original data, creating a continuously smarter system.
Proposes the 'AI-FAIR' framework to resolve fragmented data infrastructure in ecology and biodiversity, ensuring that data collectors benefit from the AI-enabled insights derived from their contributions.
Jiawen Kang, Jinchao Li, … Helen Meng
arXiv·Nov 9, 2026
By treating memory recall as a journey on a map where pauses matter, the method uses a special 'heat spread' on the map, adjusting how fast the heat spreads based on how long someone hesitates, to better understand memory problems.
Accepted at NeurIPS 2026, this spectral framework for modeling verbal recall trajectories provides a novel diagnostic tool for Alzheimer’s disease by disentangling hierarchical semantic structure from retrieval timing.
Andreas Bjerregaard, Rachael DeVries, … Frederikke Isa Marin
arXiv·Nov 2, 2026
A new set of tests helps scientists see if AI can truly design DNA by checking if it understands how tiny changes affect genes in different situations, not just if it gets big predictions right, showing it's getting smarter but still has tricky parts to learn.
Introduces a comprehensive benchmark for DNA foundation models, shifting focus from variant scoring to the critical task of guiding functional sequence design across diverse biological contexts.
Daniel P. Gilman, Jo Bovy
arXiv·Nov 6, 2026
Imagine looking through two wavy windows at two different distant lights; this paper shows that by carefully measuring how distorted the lights appear, the authors can figure out how much "extra" invisible glass is in each window, helping them better measure how fast the universe is expanding.
This paper addresses the persistent Hubble tension by utilizing double-plane gravitational lensing to break the mass-sheet degeneracy, providing a novel model-independent constraint on cosmic expansion and galaxy mass profiles.