AI and the World's Hard Problems
Science and scholarship advanced by AI, by discipline
News of AI being used to solve hard problems — in mathematics, physics, fusion, quantum science, life sciences, medicine, climate and the humanities — gathered every 3 hours from the RSS feeds and press releases of universities, research institutes, AI labs, journals and arXiv. Jev keeps only items with a real scientific result and sorts them by discipline; summaries are written only from the opening of each source — read the original for details.
RSS feed · 58 items
An interpretable hierarchical deep learning framework for lung and colon cancer detection
Classifying lung and colon cancer from histopathology images is difficult because the classes look similar and tissue morphology varies, and conventional deep learning models offer little transparency. The authors split the five-class task into coarse-to-fine stages: first normal versus cancer, then a specialized classifier for normal-tissue categories or cancer subtypes, with the best architecture chosen for each stage. The framework reached 98.50% accuracy on LC25000 and 82.01% on the independent LungHist700 dataset, and Grad-CAM, Grad-CAM++, LIME and Occlusion Sensitivity were used to highlight the image regions behind its predictions.
Artificial intelligence-based prediction of hepatocellular carcinoma in individuals without HBV or HCV infection in the MASLD era
Current liver cancer (HCC) surveillance guidelines mainly target people with cirrhosis or viral hepatitis. The researchers built an AI model to predict HCC risk in people without viral hepatitis, using data from about 7.96 million South Korean adults who underwent health screening. Among twelve models compared, MHA_DNN performed best and kept consistent accuracy in an independent validation cohort. Fatty liver index and triglycerides were the most influential predictors, suggesting support for metabolism-based personalized surveillance.
A multi task deep learning framework for automated health condition prediction from retinal fundus images using hybrid attention and explainable AI
Existing deep learning models for retinal fundus images handle only one task, offer limited interpretability, and struggle with variation across datasets. The authors propose a multi-task network that classifies diabetic retinopathy severity, glaucoma, AMD stage, and cardiovascular risk from a single image in one forward pass, with Grad-CAM explanations. It reached 94.7% accuracy and AUC 0.978 across four datasets, outperforming all single-task models statistically, and AUC of at least 0.951 in zero-shot cross-dataset evaluation.
An automata-machine learning integrated framework for intelligent mental health diagnostics
Depression among students often goes undetected until it becomes serious, so early prediction is needed. The authors combined finite automata-based behavioural modelling with machine learning, generating features from students' emotional states and merging redundant features. Using an ensemble of random forest and KStar, accuracy rose from an 82.02% baseline with random forest to 96.88% with the automata-based features. The authors state that the approach keeps the features interpretable and traceable.
LitGene: an interpretable transformer model integrating text and ontology for gene representation learning
Gene representation learning often overlooks the scientific literature, which can bias results against understudied genes. LitGene is an interpretable transformer that integrates text with Gene Ontology annotations through contrastive learning. It achieved competitive accuracy on eight protein benchmarks, zero-shot predicted GO terms, KEGG pathways and gene-disease links, and flagged ANGPTL6 as a candidate gene for idiopathic pulmonary fibrosis, supported by differential expression in independent samples.
A multi-stage validation framework for trustworthy large-scale clinical information extraction using large language models
Validating large language models for clinical information extraction usually requires costly manual annotation. The authors proposed a multi-stage framework combining rule-based filtering, semantic grounding, judge-LLM evaluation, expert review, and predictive validity analysis. Applied to substance use disorder diagnoses across 919,783 clinical notes, it removed 14.59% of LLM-positive extractions as unsupported or implausible. The primary LLM achieved an F1 score of 0.80 against judge evaluations, and its extracted diagnoses predicted later specialty care engagement better than structured-data baselines.
History’s long-secret codes are being cracked by amateurs using AI
Non-experts are increasingly using large language models to break historical ciphers, drawing on their ability to run decryption algorithms and search the internet for documents. Carter Leffen, who has no cryptography experience, asked GPT-6 Astra to decode a German army message encrypted 85 years ago, and it produced a solution by morning. Cryptography specialists welcome the interest but are overwhelmed by requests to validate AI solutions, many of which are incorrect.
Multi-level malware subfamily attribution from RGB executable images using attention-augmented deep learning
Malware analysis requires decisions at several levels, from detection to family and fine-grained subfamily attribution. The authors propose a hierarchical deep learning framework that takes RGB images of Windows PE files and performs detection, family classification and 33-subfamily attribution in one multi-output architecture. Four attention-augmented transfer learning models reached 99% binary, 96–97% family-level and 91–95% subfamily-level accuracy, with macro F1-scores up to 0.89 for fine-grained attribution.
Grounding AI agents in biological reality: Building blocks for autonomous wet labs
Frontier AI models do not reliably reason about 3D geometry and physical lab state, which makes automating liquid-handling robots difficult. Google Research grounded AI agents in physical lab robotics so they can run fast closed-loop molecular design cycles. The agents discovered protein binders with single-digit nanomolar affinity for cancer targets.
AI detects coronary heart disease from face and tongue photos in small test
A new study from China trained an AI system on photographs of the face and tongue to detect coronary heart disease (CHD). In a small test, the system correctly identified every patient with CHD.
Identification of adventitious lung sounds in children with respiratory illnesses in Bangladesh using artificial intelligence and digital auscultation
Integrated management of childhood illness guidelines have high sensitivity but low specificity for pneumonia. Health workers in Bangladesh recorded lung sounds from children under five with suspected pneumonia using a digital stethoscope, and an AI algorithm was compared with a pediatrician panel. The AI detected abnormal sounds with 61.8% sensitivity and 60.7% specificity overall, and 63.5% and 66.8% in pneumonia cases under IMCI. The authors describe this as moderate performance, with improvement needed in classifying uninterpretable recordings.
Autonomous labeling of surgical resection margins using a foundation model
Assessing surgical resection margins in pathology is affected by manual inking errors, tissue distortion, and cautery damage. The authors developed a virtual inking network that uses a frozen foundation model and a small classifier to map the margin on whole-slide images without physical ink. On 20 slides from unseen tissue blocks, it reached about 73.3% region-level accuracy and produced overlays that qualitatively matched expert annotations. The results are proof-of-concept and need validation in larger, clinical cohorts.
A machine learning-based platform for predicting the potential energy of small molecules
Accurate estimates of small-molecule potential energy after energy minimization matter for designing drug and gene-delivery molecules. The authors built deep learning models, including CNN-BiLSTM architectures with and without multihead attention, and trained them on a dataset of small-molecule structures using atom-wise sequential representations. The models reportedly outperform traditional descriptor-based machine learning methods, and the attention mechanism is said to improve generalization.
Using electronic structure to guide AI in designing new materials
Structure determines function, so scientists are increasingly using AI to design the arrangement of atoms in new materials with desired properties. Researchers at the University of Tokyo's Institute of Industrial Science extended a leading generative AI model so that it designs a material's electronic structure together with its crystal structure. The new model, MatterGen-e⁻, is reported in npj Computational Materials.
Smartphone chatting and payment habits help AI identify daily stress, small study finds
Clues to stress can be hidden in everyday smartphone habits such as chatting, searching and making payments. KAIST researchers developed a technology that analyzes several routine behaviors repeatedly performed within apps to identify a person's daily stress state.
Spatiotemporal AI-based modeling of gene regulation and variant effects in the developing human brain using Seal
The study addresses how genetic variation shapes gene regulation in the developing human brain, where data are limited. The authors built Seal, an interpretable AI transfer learning framework that predicts gene expression and variant effects across 26 regions, 30 cell types and 7 developmental stages. In autism whole-genome data, Seal found a significant burden of de novo regulatory variants in transient fetal excitatory neurons.
Forecasting Tipping Points with AI
Tipping points, after which a system cannot return to its original state, are harder to predict than short-term weather changes. A Caltech-led team showed that a recurrent neural operator (RNO), a type of neural network, can predict when tipping points occur in simple models. These include a climate-change system model and simulations of turbulent airflow over a plane's wing. The paper was published in PNAS.
Hybrid modeling for improved structure prediction of large protein complexes
The study aims to predict the structures of large protein complexes. The authors developed HDOCK-Multimer (HDM), an automated framework that combines ab initio docking with combinatorial assembly. Across three benchmarks, HDM was more accurate and robust than state-of-the-art assembly and deep learning-based methods for large protein assemblies.
AI model decodes the language cells use to communicate
During embryonic development, cells receive chemical signals that switch genes on or off and determine their roles. Tracing the sequence of signals a cell has received has been hard, because signaling pathways were assumed to differ widely across cell types. MIT and Whitehead Institute researchers developed a computational framework to address this, which could help accelerate stem cell engineering, improve disease models, and inform cancer treatments.
Advancing subseasonal forecasting with machine learning
Forecast skill drops sharply at subseasonal timescales of 2 to 6 weeks ahead. The authors introduce probabilistic bias correction, a machine learning framework that learns to correct historical probabilistic forecasts, and apply it to ECMWF's leading dynamical and AI models. It doubled the modest subseasonal skill of the AI Forecasting System and improved the dynamical model for most pressure, temperature and precipitation targets, and in ECMWF's 2025 real-time competition its global forecasts were the best.
Artificial intelligence-enabled clinical phenotyping of human brown fat
Methods for quantifying brown adipose tissue (BAT) activity in humans at scale have been lacking. Using data from 101,243 patients at a large cancer centre, the authors semi-automatically quantified BAT activity in about 8,700 patients with machine learning. Higher BAT activity, especially in the axillary depot, was associated with lower odds of cardiometabolic disease and healthier white fat distribution.
Evidence for multi-Li + clustering in lithium-ion battery electrolytes
The bulk solvation structure of lithium-ion battery electrolytes is hard to determine at atomic detail by experiment alone. The authors combined machine-learning-enabled atomistic simulations, NMR spectroscopy, electrochemical data and quantum-chemical calculations. They identified previously unstudied multi-Li+ complexes without bound anions that are easily reduced, which helps explain why empirically chosen ethylene-carbonate solvent mixtures work.
How artificial neural networks build their own learning machinery
A group of randomly connected units can reorganize into an effective information-processing network when given simple rules for when and how connections should change. The MIT team argues that an artificial network's ability to organize itself, without a developer's guidance, is a key test of biological plausibility. The researchers caution that such networks may not offer much insight into biological learning.
AI edges closer to solving the Riemann hypothesis, one of math’s hardest problems
Claude discovers a novel enzyme system
Anthropic's new life sciences lab uses Claude to explore DNA datasets for uncharacterized protein families. In early results, with only high-level direction from scientists, Claude agents found a novel enzyme system with unknown function and properties reminiscent of CRISPR. The work aims to generate hypotheses at scale and test them in the lab.
NiCLIP: neuroimaging contrastive language-image pretraining model for predicting text from brain activation images
NiCLIP is an AI model that decodes cognitive tasks, concepts and domains from brain activation patterns. It was trained on over 23,000 articles and reportedly outperforms traditional decoders at group-level decoding. It predicted cognitive tasks accurately from Human Connectome Project group-level maps, but showed limitations with noisy subject-level maps.
An all-in-one foundation model for multimodal medical image restoration and enhancement
AMIFound is a foundation model that restores and enhances medical images degraded by noise, motion artefacts, low contrast or low-dose acquisition across seven modalities, including CT, MRI, PET, endoscopy, ultrasound, fundus and X-ray. Trained on 3.76 million paired images, it outperformed 38 task-specific and general baselines on image quality metrics, clinician assessment and robustness tests. Enhanced images also improved 44 disease detection tasks, tumour stage prediction and survival risk stratification in non-small-cell lung cancer.
HIP: Hessian Interatomic Potentials without derivatives
Molecular Hessians, the second derivatives of potential energy, are needed in many computational chemistry workflows but are expensive to compute and scale poorly with system size. The authors present HIP, a deep learning model that predicts Hessians directly, without automatic differentiation or finite differences, using SE(3)-equivariant features from a graph neural network. HIP is reported to be one to two orders of magnitude faster and more accurate than automatic-differentiation Hessians, and to perform well in transition-state search, geometry optimization and vibrational analysis.
Hidden signals in sleep can predict future health
AI tools can extract physiological patterns from a single night of sleep that conventional sleep measures miss, and these patterns are linked to disease and mortality. Two recent papers, in Nature Communications and Nature Medicine, suggest sleep data hold far more health information than clinicians and researchers currently extract. Growing evidence also indicates that sleep changes can appear before a disease is diagnosed.
Machine learning analysis of cilia driven particle transport distinguishes primary ciliary dyskinesia from normal cilia
Primary ciliary dyskinesia (PCD) is a genetic condition with dysmotile cilia, and diagnosis remains challenging. The authors developed CiliaTracks, a machine learning framework that analyses particle transport on cultured airway cells captured by high-speed microscopy. Classical machine learning models and a convolutional neural network classified recordings from 14 PCD individuals and 10 healthy donors with 95-97% accuracy.
Less can be more for predicting properties with large language models
Predicting properties from coordinate-category data, such as formation energies or elastic moduli from crystal structures, is central to computational science, and large language models (LLMs) have been applied to it. Using synthetic datasets and the MatText benchmark, the authors find that LLMs consistently fail to capture coordinate information while handling category patterns well. This holds across model sizes up to 70 billion parameters, datasets up to 2 × 10^6 structures and different text representations, and for property prediction dominated by structural effects, specialized geometric architectures outperform LLMs.
ETSAM: effectively segmenting cell membranes in cryo-electron tomograms
Cell membranes are hard to distinguish in noisy cryo-electron tomography volumes. The authors present ETSAM, a two-stage method that fine-tunes the Segment Anything Model 2 (SAM2) to segment membranes, trained on 83 experimental and 28 simulated tomograms. On 17 experimental test tomograms, it outperformed other deep learning methods with high sensitivity and better precision.
Photonic circuits become trainable AI hardware
The authors demonstrate a physics-based in situ learning method that directly measures optical gradients in photonic integrated circuits. It allows diverse circuit topologies to be trained after fabrication, without requiring a full digital model of the circuit. The hardware is reported to achieve matrix design, multiwavelength computation and rapid meta-learning.
Tangermeme: a toolkit for understanding cis-regulatory logic using deep learning models
Deep learning models can predict genomic data well, but the harder step is using them for biological discovery. This paper describes tangermeme, a highly optimized toolkit for the parts of genomic deep learning other than the model itself. It shows how tangermeme can distill the cis-regulatory patterns learned by models into human-interpretable insights.
OpenAI posts 700 maths preprints online: mathematicians are up in arms
OpenAI used an unreleased model on hundreds of high-profile mathematical problems and posted more than 700 AI-generated preprints on GitHub on 6 October. Some mathematicians welcomed the solutions of long-standing problems, while others complained on social media about being scooped. One mathematician called the flood a "slopocalypse", although the content could still prove formally correct.
Generalizable perturbation prediction
Cell perturbation data reveal causal regulatory relationships, but such data are often scarce, which motivates methods that predict responses in one context from others. The authors developed State, a machine-learning model that combines a transformer-based state transition model for single-cell perturbation effects with a pretrained cell-embedding model. It was trained on perturbation data from over 100 million cells and observational data from 167 million cells, and the team also built Cell-Eval, an evaluation framework for comparing models.
Brain tumor segmentation using particle swarm optimized histogram equalization and a VGG19 based U-Net
Brain tumor segmentation in MRI is slow, error-prone, and affected by operator bias, and deep learning models often lack interpretability. The authors propose SPOHE-VGGNet, which combines a modified U-Net with VGG-19 and particle swarm optimized histogram equalization to enhance MRI contrast. It reached Dice scores of 0.9522 (meningioma), 0.9428 (glioma) and 0.9518 (pituitary tumor) on the Figshare dataset and 0.9463 for whole-tumor segmentation on BraTS 2021, though the authors state that external, volumetric and clinician-centered validation is still required before clinical use.
Unlocking Earth AI’s planetary geospatial foundation models for global public health
Public health decisions are often slowed by reporting lags, siloed data, and sparse data. Google's Population Dynamics Foundation Model (PDFM) compresses privacy-preserving search trends, human mobility, built-environment density, and environmental factors into location embeddings. These embeddings can be plugged into existing epidemiological models without task-specific fine-tuning, and five partner case studies were presented.
Accurate and well-powered case–control analysis of spatial molecular data
Existing methods for spatial molecular data often use restrictive assumptions that can miss important disease signals. VIMA, a deep-learning method, uses an ensemble of variational autoencoders to summarize small tissue patches and identify microniches whose abundance correlates with case or control status. In simulations it was well calibrated and powerful, and applied to datasets spanning three diseases it recovered known biology and identified new spatial features of disease.
ProFormer: generalizable classification of single-cell and plasma proteomes using deep learning
Current proteomic workflows have low throughput and require extensive data processing, which slows sample classification. ProFormer is a transformer-based deep-learning pipeline that takes MS1-level peptide ion features directly from mass spectrometry data. It outperformed traditional machine learning and CNNs across 14 architectures and classified single-cell proteomes by cell type, cycle stage, and differentiation trajectory, as well as patient disease status from plasma proteomes.
CellART: a unified framework for extracting single-cell information from high-resolution spatial transcriptomics
High-resolution spatial transcriptomics platforms either capture sparse transcript counts per spot or measure only a limited number of genes, which makes single-cell information hard to extract. CellART combines staining images, spatial transcriptomics data and single-cell RNA-seq references, using deep learning and probabilistic modelling to segment cells and annotate cell types simultaneously. The authors report that it is efficient, generalizable and robust, and that its outputs are compatible with widely used community tools.
A deep learning-driven pipeline for differentiating hypertrophic cardiomyopathy from cardiac amyloidosis using 2D multi-view echocardiography
Hypertrophic cardiomyopathy (HCM) and cardiac amyloidosis (CA) look alike on echocardiography, which makes them hard to tell apart. The study developed a deep-learning framework that classifies 2D echo images into five clinically relevant views, extracts features from each view and fuses them. In fivefold cross-validation on 212 HCM, 119 CA and 200 control patients, it reached a precision of 0.83, sensitivity of 0.81, specificity of 0.89 and micro-F1 of 0.82, and the authors describe it as a reliable diagnostic tool.
This light-powered AI can spot deepfakes with nearly 98% accuracy
Researchers at UCLA built an optical-neural processor that uses light to detect deepfake videos, aiming to screen the growing volume of AI-generated video. Unlike conventional systems that examine videos one by one, it analyzes 15 or more video streams at once during a single optical pass. The team reports nearly 98% detection accuracy, and the study was published in eLight.
Claude-shaped science
Prof. Matthew Schwartz describes how, after he stopped fighting Claude and let it find problems suited to its current capabilities, he built BootLoops, a toolkit for exact calculations in quantitative science. Because similar calculations recur across fields, Claude found connections to ecology, population genetics, and a dozen other areas, though these were often technically correct but initially unremarkable. Schwartz then worked with domain experts to steer BootLoops toward questions those fields care about.
Large language models discover complementary heuristics for combinatorial optimization
Combinatorial optimization underpins decisions in manufacturing, logistics and energy, but designing heuristics for each new variant takes expert effort over months, and single-pass LLM code generation often fails. LACE splits algorithm design into schemas, a tool library and a heuristic portfolio, then evolves complementary specialist heuristics within runtime limits. Across 36 classical CO-Bench problems it averaged 0.945, versus 0.870 for the strongest prior LLM-based method and 0.571 for direct prompting.
This game-playing AI is the new champ at Stratego
Researchers built an AI system for Stratego, a two-player board game in which the opponent's pieces are hidden and which is often used to test strategic thinking. Combining efficient training algorithms with new techniques for decisions under hidden information, the system defeated top-ranked human players by a large margin and outperformed the next best models while being cheaper to train. It also beat top human players in other strategy games with different rules.
Towards foundation-style models for energy-frontier heterogeneous neutrino detectors via self-supervised pretraining
Event interpretation in energy-frontier neutrino detectors is impractical for conventional reconstruction and hard for supervised models when labelled data are scarce. The authors pretrain a sparse Vision Transformer with masked-autoencoder reconstruction and relational voxel-level objectives, then fine-tune it jointly on classification and regression. On simulated events from the proposed FASER calorimeter, pretraining improved flavour and charm-quark identification, momentum regression and vertex reconstruction, and with about 1,000 labelled events the pretrained encoder matched the flavour-classification performance of a model trained from scratch.
Forecasting space weather risks on power grids
Extreme space-weather events can induce currents in transmission networks that damage equipment, and grid operators need to know when and where the effects will be most severe with enough warning. Microsoft Research built a machine learning pipeline that combines solar-wind data, AE and Dst forecasts, local latitude, geology, and ground conductivity to produce risk estimates for 66,935 substations in the continental United States. During the evaluation period it detected nearly 80% of major space-weather events and can warn operators 30 to 60 minutes before a specific risk appears.
D4RT: Unified, Fast 4D Scene Reconstruction & Tracking
AI tool for mapping smallholder crops proves itself in Senegal
Senegal's small, rain-fed farms leave much of the population exposed to climate shocks, yet satellite crop-mapping tools are largely out of reach. University of Cambridge researchers used Tessera, an open-source AI model trained on satellite images, to map crops in Senegal's groundnut basin. In tests it gave the correct result 84% of the time, more accurate than current methods, while using only a fraction of the computing resources and labelled data. In one scenario it performed 28% better than the next-best model.
Introducing Quine: An AI research system designed for the complexity of biology
Microsoft Research's Quine is an early-stage effort to build a multimodal world model of biology and an interactive harness that connects models, scientific tools, literature, and researchers, helping scientists prioritize hypotheses before lab testing. In collaboration with researchers at the Broad Institute of Harvard and MIT, the team used it to prioritize compounds predicted to drive therapeutic tumor-state shifts and validated several top-ranked candidates across multiple wet-lab assays. The system is experimental, intended for research only, and its outputs require review by qualified researchers.
New formulation helps RNA vaccines withstand high temperatures
RNA vaccines usually need ultracold storage, so MIT researchers used an AI algorithm to tweak the formulation around the lipid nanoparticles that deliver mRNA vaccines. The algorithm predicted optimal formulations, cutting the number of experiments needed and speeding development. The resulting vaccines stayed stable for up to a year at room temperature or two months at nearly 100°F, and in mice a Covid-19 vaccine produced an immune response as strong as that of an RNA vaccine similar to Moderna's.
Regional climate risk assessment from climate models using probabilistic machine learning
Climate models are too coarse for regional risk decisions, and existing downscaling methods are too costly for large ensembles, miss multivariate dependencies, or need time-aligned training pairs that free-running projections lack. GenFocal is a generative AI framework that produces statistically accurate fine-scale weather from coarse projections without paired training data. It synthesises long-lived hazards such as heatwaves and tropical cyclones, samples rare high-impact events more accurately than leading methods, and improves regional estimates of compound extremes.
Claude has improved on a longstanding lower bound for the fraction of zeros of the Riemann zeta function that satisfy the Riemann hypothesis
The Riemann hypothesis, an unsolved problem, states that the zeros of the Riemann zeta function all lie on a certain vertical line. While attempting the problem, an unreleased research version of Claude did not prove it but improved a longstanding lower bound on the fraction of zeros that satisfy it, raising it from 41.6% to 67.2%. Two Anthropic mathematicians validated the paper and wrote an informal expert note, and Claude also produced a formally verifiable proof. The result is not expected to lead to a proof of the hypothesis itself.
Claude computes a nine-loop amplitude in N=4 super-Yang-Mills
Physicist and science writer Matt von Hippel challenged AI companies to solve a hard theoretical physics problem: computing a nine-loop amplitude in N=4 super-Yang-Mills. Claude computed it, beating the challenge about a month later.
Estimating suicide risk from text
Identifying people at high risk of suicide is a priority for counselors during mental health crises. Researchers at MIT's McGovern Institute developed a language-processing tool that searches text for words and phrases linked to 49 suicide risk factors to estimate an individual's risk. In text conversations with crisis counselors, the tool accurately predicted suicide risk, and with more validation it could support risk assessment in clinical and crisis-support settings.
MIT’s tiny flying robot gets 450% faster with AI
Insect-scale flying robots have typically been slow and limited to simple flight paths. MIT researchers developed an AI-based controller with a two-part control system that enables repeated body flips. Compared with the team's previous best results, the robot's speed rose about 450% and its acceleration about 250%, and it completed 10 somersaults in 11 seconds even under wind disturbances.
How Claude is uplifting biomolecular modeling
Open-source models that scientists use to predict and design biomolecules were slow and memory-hungry. Working within Claude Science, Claude optimized more than 30 of these models in just under four weeks, speeding them up roughly 4x on average. It also built a low-memory mode that allows accurate prediction of systems larger than 10,000 tokens on a single NVIDIA GPU node, and the optimized code is being open-sourced.