Finally, huge thanks to the organizers for giving us this opportunity @algonautsproject.bsky.social

What's next? We plan to publish a more in-depth analysis of VIBE’s internal dynamics and feature–parcel mappings, to advance our understanding of brain function, and guide future work in neuroscience.
Spoiler: We already have a model that beats the top score of this challenge ;)

Shapley (MSA) Insights:
Feature attribution maps align with neuroanatomy, but crazily enough, textual features from the transcripts are the most predictive of all.

We have strong lifts over baseline both in-distribution and OOD. VIBE (final): r=0.3225 (ID, Friends S07) & 0.2125 (6×OOD).
Competition submission (earlier iteration): r=0.3198 ID / 0.2096 OOD → 1st in Phase-1, 2nd overall.
vs baseline 0.2033/0.0895 → +0.119/+0.123.

We present VIBE: Video-Input Brain Encoder, which is a 2-stage Transformer. First stage merges text, audio, and visual features per timepoint (plus subject embeddings). And the second stage models temporal dynamics with rotary positional embeddings.

Competition & Data: Algonauts 2025 tests how well we can predict brain activity while people watch naturalistic movies. Multi-modal stimuli (video, audio, text) → whole-brain fMRI, split into parcels. Train on movies/TV; evaluate in-distribution and on out-of-distribution films.

I personally prefer subway surfers. This one couldn't hold my attention for the whole video

Aside from a few mispronunciations, the AI really got the paper. Honestly, it's reassuring—if an AI can follow it without any mistakes, then folks in the field probably can too :)

Not really. Thankfully, Max Planck has GPU clusters that I can use.
Although I did ask my friend (o3) about it, according to whom, 5090 is sufficient for most cases. (chatgpt.com/share/685c6a...)