(Press-News.org) Engineers at the University of California San Diego have created a new method to make large language models (LLMs) — such as the ones that power chatbots and protein sequencing tools — learn new tasks using significantly less data and computing power.
LLMs are made up of billions of parameters that determine how they process information. Traditional fine-tuning methods adjust all of these parameters, which can be costly and prone to overfitting — when a model memorizes patterns instead of truly understanding them, causing it to perform poorly on new examples.
The new method developed by UC San Diego engineers takes a smarter approach. Instead of retraining an entire model from scratch, it updates only the parts that matter most. As a result, the new method cuts costs and is more flexible and better at generalizing what it learns compared to existing fine-tuning methods.
The researchers showed that their method can fine-tune protein language models — which are used to study and predict the properties of proteins — even when very little training data are available. For example, in predicting whether certain peptides can cross the blood-brain barrier, the new method achieved higher accuracy than conventional methods while using 326 times fewer parameters. In predicting protein thermostability, it matched the performance of full fine-tuning while using 408 times fewer parameters.
“With our method, even small labs and startups without huge budgets, supercomputer-level resources or large datasets can adapt large AI models for their own needs,” said Pengtao Xie, a professor in the Department of Electrical and Computer Engineering at the UC San Diego Jacobs School of Engineering. “This work represents a step toward democratizing AI.”
The new method for fine-tuning and adapting LLMs was published in Transactions on Machine Learning Research. This research was supported by the National Science Foundation (IIS2405974 and IIS2339216) and National Institutes of Health (R35GM157217 and R21GM154171).
END
AI models can now be customized with far less data and computing power
2025-10-21
ELSE PRESS RELEASES FROM THIS DATE:
Twenty-five centers join Bronchiectasis and NTM Care Center Network
2025-10-21
Miami (October 21, 2025) – The Bronchiectasis and NTM Association has accepted eight Care Center and 17 Clinical Associate Center sites in 14 states into the Bronchiectasis and NTM Care Center Network (CCN). The CCN includes 58 centers across the United States.
The CCN aims to facilitate access to specialized care and support for the hundreds of thousands of people with bronchiectasis and nontuberculous mycobacterial (NTM) lung disease.
“The prevalence of bronchiectasis and NTM lung disease continues to increase. Patients deserve access to high-quality, specialized care and resources,” said Doreen Addrizzo-Harris, M.D., Chair of ...
Botox-like substance brings relief to Ukrainian war amputees
2025-10-21
Study involved 160 amputees treated at two hospitals in Ukraine
At one month, botulinum toxin group saw a four-point pain drop versus one point for standard care group
At three months, the trend shifted as effects of botulinum toxin waned
Senior author is a retired U.S. Army colonel and physician who traveled to Ukraine to launch the study and collaborate with local doctors
CHICAGO --- Botulinum toxin injections provided greater short-term relief for phantom limb pain than standard medical and surgical care among Ukrainian war amputees, reports a new study led by Northwestern ...
People with dark personality traits use touch to manipulate their partners
2025-10-21
A hug can soothe your mind, reduce your stress and actually activate oxytocin, the “love hormone,” in your body. But new research from Binghamton University, State University of New York reveals that not all hugs are harmless – some partners use touch as a means of control.
People with “dark triad” personality traits – narcissism, psychopathy and Machiavellianism – are more likely to use touch to manipulate their partners, according to a new paper published in Current Psychology by Richard Mattson, professor of psychology at Binghamton University, and a team of students.
“What’s new about our work ...
It’s not just diet: where a child lives also raises type 2 diabetes risk
2025-10-21
Type 2 diabetes (T2D), once considered an adult-onset disease, is increasing at alarming rates in children and adolescents. Before the mid-1990s, just 1% to 2% of youth with diabetes had T2D. Today, that number has skyrocketed to between 24% and 45%, with the average age of diagnosis hovering around 13 years old.
This troubling trend closely tracks with the ongoing rise in childhood obesity. While genetics, diet and physical activity all play roles in T2D risk, new research from Florida Atlantic University’s Charles E. Schmidt College of Medicine highlights another key factor in T2D risk: where a child lives.
Researchers conducted a large-scale study to explore ...
Predicting physical activity change after a cardiovascular diagnosis
2025-10-21
Brain connectivity patterns and environmental factors predict which older adults will successfully increase physical activity after receiving a cardiovascular diagnosis. Nagashree Thovinakere and colleagues studied 295 cognitively healthy but physically inactive older adults from the UK Biobank who received cardiovascular diagnoses during a roughly four-year period. The authors tracked which people increased their activity level to the moderate-to-vigorous physical activity levels recommended by the World Health Organization, using both self-reports and accelerometer data. The authors used machine learning to ...
Algorithmic outreach leads to information inequality
2025-10-21
Algorithms that identify influential people in social networks can help maximize the reach of messages, but a modeling study shows that those same algorithms can disseminate information inequitably, potentially exacerbating existing social inequalities. From public health campaigns to information about social services, algorithms that identify “influencers” have been used to maximize reach. Vedran Sekara and colleagues used the independent cascade model on synthetic and diverse real-world social networks, including connections between households in multiple villages, connections between political bloggers, Facebook friendships, and scientific collaborations. The authors ...
Szeged researchers accelerate personalized medicine with AI-powered 3D cell analysis
2025-10-21
The HCS-3DX platform performs automated analysis of three-dimensional cell cultures, known as spheroids. Using AI-based image processing and sample selection, the system enables large-scale, high-precision screening of cellular models within a fraction of the usual time.
“Our goal was to create a unified platform that integrates the strengths of existing technologies and can be easily implemented in research and industry” said Ákos Diósdi, first author of the study.
According to Dr. Péter Horváth, director of the Institute of Biochemistry at the HUN-REN Biological Research Centre, Szeged and senior author of ...
Offline interactions predict voting patterns better than online networks
2025-10-21
Offline social networks, revealed by co-location data, predict US voting patterns more accurately than online social connections or residential sorting. Michele Tizzoni and colleagues analyzed large-scale data on co-location patterns from Meta’s Data for Good program, which collates anonymized data collected from people who enabled location services on the Facebook smartphone app. Colocation is defined as two people being within the same map tile, which is less than 600×600 meters, depending on latitude. The political affiliation of each person was inferred ...
Hanyang University researchers develop novel facet guided metal plating strategy, improving stability anode-free metal batteries
2025-10-21
Anode-free metal batteries represent an exciting new design, where prefabricated anodes are eliminated to maximize energy densities. For example, in magnesium (Mg) metal batteries, instead of starting with an Mg anode, only a bare metal, usually copper (Cu) or Zinc (Zn), current collector is used as the anode side. When the batteries are first charged, Mg from the cathode deposits directly onto this collector, forming a thin Mg layer that acts as the anode. This avoids excess anode materials, making batteries lighter, more compact, and cheaper. Unfortunately, these batteries suffer from dendrite formation, which significantly affects battery ...
When cells run a red light: Double trouble for old models in cell division
2025-10-21
Scientists at the Ruđer Bošković Institute (RBI) in Zagreb, Croatia, have discovered that the protein CENP-E, long believed to act as a motor dragging chromosomes into place during cell division, in fact plays a completely different role in chromosome movement. It stabilizes the first attachments of chromosomes to the cell’s internal “tracks,” ensuring they line up correctly before being divided. In a related study, scientists found that small structures inside our cells, called centromeres, which were once thought to function independently, help guide this key ...