🔬 AI Research

Academic papers, research breakthroughs, and scientific advances in machine learning and AI.

585 articles

NSF funds Texas A&M tool to speed up materials discovery

The National Science Foundation is backing a new free, open-source software project that uses machine learning to help materials scientists answer a common research question faster. The tool aims to make discovery workflows more accessible to the wider scientific community. Source: news.engineering.tamu.edu

Google unveils TimesFM-3, a zero-shot model for multivariate forecasting

Google's new TimesFM-3 is a foundation model built for time series forecasting that can handle multiple variables at once without task-specific training. It aims to deliver state-of-the-art accuracy across diverse forecasting scenarios in a single unified approach. Source: research.google

Study: combining econometrics with ML boosts volatility forecasts when capacity is controlled

New research challenges the idea that bigger models and more data automatically mean better financial predictions. Instead, combining traditional econometric methods with machine learning — while carefully controlling model capacity — improves volatility forecasting accuracy. Source: bioengineer.org

New ML tool tracks stablecoin transfers to spot money laundering

Researchers built a behavior-based AI system that analyzes USDT and USDC transactions on Ethereum, classifying wallets as sanctioned/frozen, linked to cybercrime, or legitimate. The tool aims to help authorities trace illicit stablecoin flows more effectively. Source: eurekalert.org

AI models help flag Parkinson's patients at risk of faster disease progression

Researchers at the University of Miami developed machine-learning models that combine clinical, biomarker, and imaging data to identify Parkinson's patients likely to decline more rapidly. The approach could help doctors intervene earlier with more targeted treatment. Source: newswise.com

New ML framework maps potential microbial symbiotic relationships from genome data

A machine-learning tool called symclatron analyzes genomic sequencing data to predict whether uncultivated bacteria and archaea live independently or form symbiotic relationships with other organisms. The framework offers a new way to study microbes that can't be grown in lab conditions. Source: nature.com

From MIT to IBM: speeding up AI and quantum deployment

A collaboration linking MIT research with IBM's technology stack aims to accelerate how AI and quantum computing move from lab to real-world use. Details on specific projects or timelines were not provided. Source: news.mit.edu

PACMAN: new AI framework keeps fusion reactors safe with millisecond decisions

PACMAN is an AI-based control framework designed to manage fusion systems safely, making critical decisions within milliseconds. Its speed is meant to help prevent instabilities that could disrupt or damage fusion reactors. Source: phys.org

CSIRO's Data61 creates a 'vaccine' to shield machine learning from attacks

Australian research agency CSIRO, through its Data61 unit, has developed a defense technique designed to protect machine learning models from adversarial attacks. The method works like a vaccine, exposing models to manipulated inputs in advance so they become more resistant to malicious interference. Source: zdnet.com

New Approach Combines Feature Engineering with Explainable LightGBM to Predict Concrete Strength

A study presents a method for predicting concrete strength using domain-specific feature engineering paired with the explainable machine learning model LightGBM. The approach aims to improve prediction accuracy while keeping the model's decisions interpretable for engineers. Source: nature.com

Can AI Models Spot Parkinson's Patients Facing Faster Decline?

Researchers are exploring whether AI models can identify Parkinson's patients likely to experience more rapid disease progression. Early detection of high-risk patients could help tailor treatment and monitoring strategies. Source: news.med.miami.edu

Federated learning could let AI models train without exposing private data

Researchers are exploring federated learning as a way to train large language models across distributed data sources without centralizing or sharing sensitive information. The approach could help address privacy concerns tied to training AI on private datasets. Source: bioengineer.org

Google AI unveils EnvHarness, a framework that makes agent training environments adaptive

Google AI has introduced EnvHarness, a programmable layer designed to convert static agent environments into dynamic, adaptive training worlds. The tool aims to give AI agents more realistic and flexible conditions to learn from during training. Source: marktechpost.com

New 'Code-as-World' AI Agent Turns Real Videos into Executable Physics Simulations

Researchers introduced 'Code-as-World,' an agentic system that converts real-world video footage into executable MuJoCo physics programs. The approach uses an iterative loop to translate visual scenes into code-based simulations that can be run and manipulated as physics models. Source: marktechpost.com

AI Detects Early Pregnancy Risks in Global Study of 500,000 Cases

A large-scale study analyzing data from 500,000 pregnancies worldwide found that AI models can identify risk indicators early in pregnancy. The findings suggest AI-based screening could support earlier intervention and better outcomes for mothers and babies. Source: labroots.com

Machine Learning Tracks How Immune Response Shifts During Sepsis

Researchers used machine learning to map how the immune system's response evolves over time in patients with sepsis. The approach could help identify critical shifts in immune activity that inform treatment timing. Source: news-medical.net

Agent Seer: Generating Test Scenarios from Spec Understanding

A new research approach called Agent Seer synthesizes test scenarios by having AI systems interpret and understand written specifications. The method aims to automate scenario creation for validating complex software behavior. Source: machinelearning.apple.com

Flawed benchmarks may be skewing AI drug-discovery rankings

Researchers warn that widely used benchmarks for evaluating AI models in drug discovery may contain hidden flaws, potentially inflating or misrepresenting performance results. This raises concerns about the reliability of current leaderboards used to compare AI systems in pharmaceutical research. Source: chemistryworld.com

New AI framework promises smarter control of solid oxide electrolysis cells

Researchers have developed an AI-based framework to optimize the operation of solid oxide electrolysis cells, a technology used for efficient hydrogen production. The approach aims to boost performance and stability of these energy systems. Source: techxplore.com

Machine learning estimates key phytoplankton pigments from satellite ocean-color data

A study demonstrates how machine learning can go beyond simple chlorophyll measurements to estimate diagnostic phytoplankton pigments using multispectral ocean-color satellite data. This offers a more detailed picture of marine ecosystem health. Source: astrobiology.com