Researchers at the International Institute of Information Technology (IIIT-Hyderabad), in association with peers from some other organisations, have identified a “sweet spot” in brain-artificial intelligence similarity studies.Their study reveals that a smaller language model can achieve brain alignment comparable to larger models, offering a more efficient, accessible approach to advanced AI.Smaller, efficient AI models are highly significant for India. They require less computational power, allowing researchers and developers to deploy advanced AI solutions on smaller instruments or mobile phones, overcoming limitations in access to massive cloud or GPU infrastructure.Subba Reddy Oota, Vijay Rowtula, and S Bapi Raju of IIIT-H presented the findings of the paper “Linguistic properties and model scale in brain encoding: from small to compressed language models” at the International Conference on Machine Learning (ICML) held in Seoul recently. The findings debunked the myth that larger language models understand the human brain better.By facilitating the development of Brain-Computer Interfaces (BCIs), brain decoding using these smaller models offers a significant opportunity to help individuals with neurological conditions restore their ability to communicate and move.“As newer AI algorithms are developed, we are looking at how similar, or ‘aligned’, they are to how the human brain processes information. We are evaluating the similarities between human brain processing and modern AI algorithms,” S. Bapi Raju, a Professor and Head of Cognitive Sciences Lab at IIIT Hyderabad, told businessline.Modern AI models are characterised by their complexity, which is measured in the number of parameters, such as 3 billion, 7 billion, 10 billion, or 14 billion parameters. “Generally, as the size increases, its alignment with the brain increases as well. However, we wanted to find the smallest model that still provides significant alignment, rather than relying on very large models,” he said.“We investigated a range of models from 1.5 billion up to 14 billion parameters. We found that the 3-billion parameter model is a sweet spot. It provides performance and brain alignment that is just as good as the 14-billion parameter model. However, if we drop down to a 1.5-billion parameter model, we observe a noticeable degradation in brain alignment,” he said.What is the impact of this study?Looking ahead, Raju believes this discovery will have a direct impact on two major areas of neuroscience. “We are planning to use these 3-billion parameter models to study how the brain stores information, utilising deep learning models to better understand human brain processing,” he said.It also addresses the “brain decoding problem.” For example, the brain’s response when viewing an aeroplane. How it reconstructs what the person is experiencing based purely on that brain response.“Brain decoding is highly useful for designing Brain-Computer Interfaces (BCIs). For individuals who, due to neurological disorders, have lost the ability to speak or move their limbs, we can look at their brain activity and decode it to develop appropriate BCI tools,” he said.In the long run, understanding these smaller models will be very useful for developing those interfaces.Published on July 28, 2026