The Capacity of Thought: Benchmarking Llama 3.2 in Semantic fMRI Neural Language Decoding and Improving the Huth Encoding-Model Baseline

Researchers improved the Huth et al. fMRI encoding pipeline by expanding voxel selection to 15K and substituting GPT-2 medium for GPT-1, achieving an 11% relative METEOR gain. They also introduced fMRIFlamingo, which maps BOLD activity to a frozen Llama-3.2-1B via trainable cross-attention layers. Despite scoring 42.86% Top-1 accuracy on a ranking task, a blind control ablation with zeroed fMRI inputs yielded near-identical scores, revealing that apparent decoding success stems from the language model's prior rather than neural input.