fMRIFlamingo maps BOLD activity to Llama-3.2-1B but relies on frozen prior

The article investigates semantic fMRI neural language decoding using Llama 3.2 and improves the Huth et al. encoding pipeline. It introduces fMRIFlamingo, which maps BOLD activity to a frozen Llama-3.2-1B model via a learned brain tokenizer and Perceiver Resampler.

• The improved Huth pipeline uses expanded voxel selection (10K to 15K) and GPT-2 medium, achieving mean METEOR = 0.149 and BLEU-1 = 0.200 for subject UTS03.
• fMRIFlamingo achieves 42.86% Top-1 accuracy on a 1-in-100 ranking task.
• A blind control ablation with zeroed fMRI inputs yields near-identical scores, indicating decoding success is driven by the frozen language prior rather than neural input.

These results demonstrate that high-capacity language models do not inherently improve fMRI decoding and can obscure failures without rigorous blind-control evaluation.