Improved AI Decodes Brain Activity into Words with 11% Accuracy Gain

A new study on arXiv presents two complementary investigations into decoding continuous language from fMRI signals, a core challenge in non-invasive brain-computer interface (BCI) research. The researchers improved the Huth et al. ridge regression encoding pipeline by expanding voxel selection from 10,000 to 15,000, substituting GPT-2 medium for GPT-1 as the beam-search proposal model, and using GPU-accelerated bootstrap training. These enhancements achieved a mean METEOR score of 0.149 and BLEU-1 of 0.200 across three held-out narratives for subject UTS03 — an 11% relative METEOR gain over the replication baseline.

This matters because it demonstrates a clear path to improving the accuracy of decoding thoughts from brain activity, which could eventually enable communication for people who cannot speak due to injuries or conditions like ALS. While still in early stages, the work refines the encoding-model approach and provides a stronger baseline for future research in semantic fMRI neural language decoding.

The paper is available on arXiv under the identifier 2607.12079.