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
Devin Pereira, Willem Zuidema
arXiv·Aug 7, 2026
Big AI models can 'see' the whole puzzle in their mind, but then they forget parts of it while trying to solve it, making them fail; fixing their memory helps them succeed.
Mechanistically probes whether Large Reasoning Models (LRMs) utilize emergent world models for planning, challenging the 'illusion of thinking' in the wake of recent test-time scaling breakthroughs.
Juan J. Romero, Andrew Gall, … Manuel J. Llansola-Portoles
arXiv·Aug 7, 2026
Using a super-fast camera that sees tiny vibrations, scientists finally figured out that protective carotenoid triplets in plants come from energy hand-offs from chlorophyll, not from carotenoids splitting their own light energy.
This study identifies the specific electronic pathways for carotenoid triplet formation in light-harvesting complexes, resolving a long-standing debate on photoprotection mechanisms in photosynthetic bacteria.
Joseph M. Cavanagh, Jonathan B. Arnold, … Teresa Head-Gordon
arXiv·Jul 15, 2026
By teaching powerful language models the 'language of molecular geometry' using Z-matrices and smart fine-tuning, the authors enable them to accurately predict 3D shapes of molecules from simple text descriptions.
Demonstrates that fine-tuned LLMs using Z-matrix representations can outperform specialized geometric deep learning models in predicting equilibrium structures and conformers for drug-like molecules.
Philip Arathoon, Gavin Ball, Matthew D. Kvalheim
arXiv·Jul 29, 2026
By carefully placing five tiny electric charges, the authors found a way to create at least 24 stable spots where a test charge would feel no push or pull, proving an old idea about how many such spots are possible was wrong.
Disproves a 150-year-old electrostatics conjecture using a five-charge counterexample; gained viral attention for the authors' disclosure that the successful construction was initially suggested by an LLM.
Tobias Haug, Askery Canabarro, Leandro Aolita
arXiv·Jul 21, 2026
A new machine learning trick helps quantum computers throw out 'bad' results by learning simple error patterns, making their final answers much more trustworthy without needing complex calculations.
Introduces a decoder-agnostic ML method for QEC post-selection, validated on QuEra's neutral-atom hardware to significantly improve logical fidelity in magic-state distillation and surface codes.
Fan Yang, Yuting Su
arXiv·Aug 4, 2026
A new robot brain called LiLa-WAM learns to predict what will happen next and how to move, all at once, using a compact mental picture and a special visual task cue instead of words, making it super efficient.
Introduces a lightweight latent reasoning space for World-Action Models, enabling end-to-end training on a single 24GB GPU while achieving high success rates on complex manipulation benchmarks.
William E. Chapman, John Schreck, Yingkai Sha
arXiv·Jul 20, 2026
When AI weather models are forced to perfectly conserve things like water, the way they learn can hide a big problem: the model's raw predictions can get worse, even though the final, corrected output looks perfect.
Identifies a critical 'scale degeneracy' where physical budget correctors mask model degradation during training, challenging the reliability of current AI weather emulators and proposing a robust supervised penalty approach.
Jan Linnenbrink, Jakub Nowosad, … Hanna Meyer
arXiv·Jul 22, 2026
A new tool called STeMP helps scientists clearly describe their environmental computer models and even warns them about common mistakes, making their work more trustworthy and easier for others to understand and reuse.
Introduces a standardized reporting protocol and R package for spatio-temporal machine learning in environmental research, addressing the 'black box' challenge and improving model transparency and reproducibility.
Andrew L. Smith, Linxing Preston Jiang, … Stefano Recanatesi
arXiv·Aug 6, 2026
By flipping how a learning rule works, the model makes brain-like layers quiet when things are expected and noisy when they're surprising, showing how brains might predict without complex math.
Proposes a biologically plausible 'Inverted Forward-Forward' algorithm implementing hierarchical predictive coding through local activity cancellation, bridging synaptic learning rules with global cortical computations in a new NeuroAI framework.
Aleksandr Sharipov, Yusif Mukhtarov, Igor Molybog
arXiv·Aug 3, 2026
Scientists created a new AI that turns complex cell data into simple codes, then uses these codes to create new, realistic cell data, showing that bigger models and more data predictably improve its performance.
This paper establishes the first jointly-fit two-exponent scaling law and compute-optimal frontier for single-cell foundation models, providing a critical framework for optimizing future large-scale transcriptomic generative models.
Laura Burn, Nelson Christensen, Richard Easther
arXiv·Aug 7, 2026
Scientists explored if tiny black holes, if they are dark matter, would make gravitational waves when zipping past giant black holes in galaxies, finding most signals too weak for current detectors.
This paper proposes a novel method for detecting primordial black holes as dark matter candidates by analyzing gravitational wave bursts from hyperbolic encounters with supermassive black holes.