Aligning brains into a shared space improves their alignment with large language models

Mapping neural activity from multiple individuals into a "shared response space" improves the accuracy of LLM-based brain encoding by 37%.

This study reveals that aligning electrocorticography (ECoG) data from multiple participants into a common information space significantly enhances the ability of Large Language Models (LLMs) to predict individual brain activity. By using a “shared response model” while participants listened to the same podcast, researchers achieved a 37% average improvement in encoding accuracy, with the most substantial gains appearing in areas specialized for language comprehension like the superior temporal gyrus. This approach allows for effective denoising and suggests that shared computational spaces are key to building neural encoding models that generalize across different human brains.

Link