mercurialsolo/tribe-replay-heatmaps

TRIBE Replay Heatmaps

Video-ID manifest and per-video results for the paper A global predicted-fMRI drive signal from TRIBE does not predict YouTube replay heatmaps.

Paper: https://arxiv.org/abs/2607.01400

Code: https://github.com/mercurialsolo/tribe-replay-heatmaps

Summary

We run TRIBE, the winning model of the Algonauts 2025 brain-encoding challenge combining Llama-3.2, V-JEPA2 and Wav2Vec-BERT, on 48 YouTube videos. We reduce its predicted cortical response to a per-second global field power engagement curve and test whether that curve predicts the crowd-sourced most-replayed heatmap of each video. The position-controlled partial correlation is +0.058, 95% CI -0.04 to 0.15, t=1.21 at 47 degrees of freedom, p=0.23. It is indistinguishable from zero and not above loudness or motion baselines.

Files

video_ids.txt: the 50 YouTube video IDs used in the study, one per line.

videos.csv: per-video table with columns id, youtube_url, category, duration_s, analyzed_s, n_timesteps, raw_pearson_r, partial_r_position_controlled, baseline_loudness_partial_r, baseline_motion_partial_r.

Data policy

We do not redistribute video files or fMRI data. Videos are identified only by their YouTube ID. The most-replayed heatmaps are public YouTube metadata fetched per ID. Everything needed to reproduce the analysis is derivable from the IDs plus the code.

Citation

If you use this manifest or the results, please cite:

@misc{sahu2026tribereplay,
  title         = {A global predicted-fMRI drive signal from TRIBE does not predict YouTube replay heatmaps},
  author        = {Sahu, Barada and Pandey, Shivesh},
  year          = {2026},
  eprint        = {2607.01400},
  archivePrefix = {arXiv},
  primaryClass  = {cs.SE},
  url           = {https://arxiv.org/abs/2607.01400}
}

Please also cite the TRIBE model.