TRIBE brain model fails to predict YouTube replay heatmaps

A new arXiv paper tests whether brain-encoding models can predict behavioral engagement, using TRIBE—the winning model of the 2025 Algonauts challenge, which combines Llama-3.2, V-JEPA2, and Wav2Vec-BERT. The researchers applied TRIBE to 48 YouTube videos, extracting a predicted fMRI global field power signal as a per-second engagement curve. They then correlated this curve with YouTube's 'most replayed' heatmap, a passive proxy for viewer re-watch behavior. If brain signals forecast what keeps audiences hooked, high correlations would be expected.

Instead, the correlation was essentially zero: a pooled position-controlled partial correlation of +0.058 (95% CI [-0.04, 0.15]; p=0.23). This result was indistinguishable from simple loudness and motion baselines (loudness +0.04, paired p=0.74). The moderate correlations previously reported for music videos turned out to be an artifact of the intro/onset-replay pattern—not genuine content prediction. The null result persisted across six different cortical-network readouts and was confirmed by autocorrelation-preserving permutation tests. The authors release code and video IDs, providing a rigorous benchmark for future work linking neural signals to real-world viewing behavior.