Milos Suvakovic, Dom Marhoefer, Glenn Grant-Richards +1 more
Reading language straight out of a brain scan sounds like science fiction, and this paper is partly a reality check on how real that reading is. The authors ran two experiments. First they tuned an older fMRI decoding pipeline, swapping GPT-2 medium in for GPT-1, which earned a small but real gain on held-out stories for one subject. Then they built fMRIFlamingo, feeding brain activity into a frozen Llama-3.2 model.
The honest twist is the point. Their fancier model picked the right sentence out of 100 well above chance, but feeding it blank, zeroed-out brain data barely changed the score. So the language model's own guessing, not the neural signal, did most of the work. A sharp reminder that a big model can hide the fact that it isn't actually listening to the brain.
This rests on the abstract, so read the paper for exact numbers and the blind-control setup.
Decoding continuous language from fMRI signals remains a core challenge in non-invasive brain-computer interface research. We present two complementary investigations. First, we improve the Huth et al. ridge regression encoding pipeline through expanded voxel selection (10K->15K), substitution of GPT-2 medium for GPT-1 as the beam-search proposal model, and GPU-accelerated bootstrap training, achieving mean METEOR = 0.149 and BLEU-1 = 0.200 across three held-out narratives…
EEG-Based Imagined Speech Decoding Using a Hybrid CNN-SNN Architecture
arXiv (neural decoding) · July 3, 2026Rethinking Brain Decoding with CLIP: The Role of Adversarial Robustness
arXiv (BCI) · July 3, 2026Stacked LoRA for Subject-Adaptive EEG Foundation Models in Motor Imagery Decoding
arXiv (BCI) · July 2, 2026Antenna System for Simultaneous Wireless Power and Information Transfer to Brain Implants