Verified by 
Artificial Intelligence
Large Language Models 2024 Year in Review and 2025 Trends
A look back at LLM milestones and a sneak peek at the year ahead.
Posted January 1, 2025 Reviewed by Monica Vilhauer Ph.D.
Exploring uses for artificial intelligence (AI) machine learning is de rigueur, and large language models (LLMs) are having their moment in the spotlight. Last year, LLMs were used in a variety of research studies to unravel complexities in psychology, neuroscience, and many other fields.
Evaluating LLM transformers using principles drawn from human cognition and psychology will be a growth area this year. Researchers have just scratched the surface in evaluating artificial intelligence LLMs using what scientists know from psychology and human behavior. For example, last year an AI study examined the rationality of large language models using cognitive psychology from a series of tasks designed to spot human heuristics and biases.
Understanding exactly how the human brain processes speech and language can accelerate advancements in a wide variety of real-world uses, such as brain-computer interfaces, AI machine learning, robotics, neurotechnology, child development, talk therapy, speech synthesizers, natural language processing, assistive technology for the disabled, commercial chatbots, voice-recognition products, machine translation, autonomous vehicles, psychology, psychiatry, medical diagnostics, pharmaceuticals, biotechnology, and health care.
In 2025, expect more exploration of conversational AI in the fields of human speech and language. Last year inroads were made using LLMs to shed light on the biological brain. For example, a 2024 PNAS study by Goldin-Meadow and others used an LLM created by Google called BERT (Bidirectional Encoder Representations from Transformers) to determine the timing of the pediatric milestone of linguistic productivity, the ability to produce novel language versus mimicking speech. This discovery has the potential to help clinicians in child development, speech therapy, and artificial intelligence development.
Another trend to watch in 2025 is the use of LLMs to identify patterns from reams of complex biological data captured by established brain activity imaging and recording technologies. Expect an increase of research studies using LLMs and neural activity data overall. The largest increase is expected in studies that use neural data recorded through noninvasive methods by functional Magnetic Resonance Imaging (fMRI), magnetoencephalography (MEG) and electroencephalography (EEG), or even via a temporary digital electronic tattoo (e-tattoo) printed directly on the scalp as demonstrated in 2024.
Another 2024 study published in Nature Machine Intelligence compared a variety of LLMs and actual human brain activity recordings to discover that LLMs and human brains are converging. Namely, the study showed that the LLMs studied showed hierarchical processing similar to the regions of the biological brain responsible for sound and language processing.
An emerging trend in the position to expand in 2025 is the evaluation of the predictive capabilities of LLMs versus human experts in neuroscience. Last year, an AI milestone was established with the demonstration that the predictive capabilities of large language models outperformed human neuroscience experts in determining neuroscience outcomes in the landmark study published in Nature Human Behaviour.
Also in 2024, OpenAI’s LLM GPT-4 was put to the test vis-à-vis human neuroradiologists in a study published in European Radiology in the task of diagnosing real-world clinical MRI reports of brain tumor cases. The scientists found that the performance of GPT-4 could serve as a neuroradiology advisory tool for clinicians as well as a useful second opinion on final neuroradiology diagnoses.
Beyond the realm of neuroscience, expect more integration of LLMs as potential enhancements for consumer-facing products, especially in transportation. For example, in 2024 Purdue University researchers conducted a study featuring a conversational AI based on an LLM framework called Talk2Drive that can interpret human voice commands to guide autonomous vehicles. The research was the first-of-its-kind to conduct a multi-scenario field experiment that deploys LLMs on a real-world self-driving car.
An important trend to watch in 2025 is the potential increased use of LLMs by scientists as a research tool itself throughout the research study lifecycle for a variety of functions, such as information gathering, editing, formatting, grammar check, synonym searches, background data gathering, published research searches, database cleaning, data analysis, statistical reporting, writing, rewriting, brainstorming, synthetic data production, data label generation, and more uses.
In this past year, one preprint conducted by researchers at the Allen Institute for Artificial Intelligence (Ai2), University of Washington, University of Copenhagen, and Princeton University shows that a majority (80.9%) of over 800 verified published authors listed on Semantic Scholar surveyed had self-reporting using LLMs in one or more areas in their research.
Expect this number to rise as more researchers use LLMs as an assistive technology throughout the scientific research process; from brainstorming ideas to helping with report writing. Another trend to watch is the growing call for transparency and disclosure when LLMs are used and how they are used to produce research studies in the future.
This past year was a remarkable year of discovery at the intersection of artificial intelligence transformer technology and the sciences. Expect more research studies evaluating the capabilities of large language models versus human experts, as well as the increased use of LLMs to produce the actual research itself. Using human behavior and applied psychology to understand the efficacies and capabilities of large language models will be an important growth area in 2025 and beyond.
Copyright © 2025 Cami Rosso All rights reserved.
