A Global Predicted-fMRI Drive Signal from TRIBE Does Not Predict YouTube Replay Heatmaps

Jul 16, 2026

Deep multimodal brain-encoding models predict fMRI responses to naturalistic video with high accuracy, but their ability to forecast behavioral engagement remains unclear. Using TRIBE and reducing predicted cortical responses to a per-second engagement curve (global field power) across 48 YouTube videos, the curve shows no relationship to viewers’ “most replayed” heatmaps. The null persists across confound controls, cortical-network readouts, ROI classes, and an equivalence-and-Bayes-factor framework. Additional probing suggests any weak signal is limited to the visual input stream.

LanguageEnglish

CategoryResearch & Report

FormatPDF

Pages12

Reading Time30 min

Introduction

Related Work

arXiv :2607 .0 1400v2 [ cs . SE] 4 Jul 2026 A global predicted-fMRI drive signal from TRIBE does not predict YouTube replay heatmaps Barada Sahu Shivesh Pandey Cabal AI Para AI Deep multimodal brain-encoding models now predict fMRI responses to naturalistic video with high accuracy. Whether their predicted neural signals also forecast behavioral engagement is unknown. We run TRIBE, the winning model of the 2025 Algonauts brain-encoding challenge (Llama-3.2 + V-JEPA 2 + Wav2Vec-BERT), on 48 YouTube videos and reduce its predicted cortical response to a per-second engagement curve, the global field power. Correlated against each video’s “most replayed” heatmap, a passively-collected proxy for which moments viewers return to, the curve shows no evidence of predicting re-watch behavior. The pooled position-controlled partial correlation is +0 .058 (95% CI [−0 .04 , 0. 15]; one-sample 􀁴(47) = 1 .21 , 􀁰 = 0 .23), indistinguishable from zero and not significantly above simple loudness and motion baselines (loudness +0 .04, paired 􀁰 = 0 .74) . The raw correlation is also near zero; the moderate values reported for music videos are a genre-specific intro/onset-replay artifact and do not generalize to other content. The null holds across six cortical-network readouts, signed value/salience ROIs, and an autocorrelation-preserving permutation test; a supervised leave-one-video-out probe on the predicted cortex appears to reach 􀁲 = 0 .47 but this collapses to a shared temporal-shape artifact under a proper position control. Running the same probe on TRIBE’s three input streams reveals at most a small, borderline content-specific signal in the visual stream (matched vs. mismatched 􀁰 ≈ 0.004–0.06 across feature extractions) and none in audio, text, or the predicted cortex, tentatively placing what little signal exists in the visual input, upstream of the encoding. The inter-subject-correlation (ISC) readout, the closest prior positive result at this grain, is unavailable from the subject-averaged released model, so we fit our own per-subject encoders on the Algonauts fMRI (validated in-domain at 􀁲 ≈ 0.15 and cross-domain, Friends→film, at 􀁲 ≈ 0. 10); the predicted ISC still does not track re-watch (􀁲 = −0 .04 , 􀁰 = 0 .34) . We bound the claim rather than merely fail to reject it: a Bayes factor gives moderate evidence for the null (BF01 = 3 .2), an equivalence test excludes effects larger than 􀁲 ≈ 0. 14, and the target’s split-half reliability (≈ 0 .82; ceiling 􀁲 ≈ 0 .9) shows the null is not a noisy-label artifact. We release the code, the video-ID manifest, and an acquisition method that works despite YouTube’s SABR-only streaming. Date: July 7, 2026 Correspondence: [barada@gmail.com](barada@gmail.com), [cs21bt067.alum25@iitdh.ac.in](cs21bt067.alum25@iitdh.ac.in) 1 Introduction Encoding models that predict brain activity from naturalistic stimuli have improved sharply, with deep multimodal architectures such as TRIBE winning the 2025 Algonauts challenge (out of 263 teams) by mapping fused text, video, and audio features onto the cortical surface (d’Ascoli et al. , 2025) . Separately, the neuroforecasting literature shows that measured neural signals (fMRI/EEG) can predict aggregate population behavior beyond self-report, from cultural popularity (Berns and Moore, 2012) to crowdfunding and market outcomes (Genevsky et al. , 2017), and that the temporal reliability of neural processing tracks audience preferences (Dmochowski et al. , 2014 ; Hasson et al. , 2004) . Whether predicted neural signals, which require no scanner and are inexpensive to compute, inherit this predictive power has not been tested. The expected direction is not obvious. An accurate encoder might preserve the behaviorally-relevant structure of the measured response; it might equally regress that structure toward the group mean, discarding exactly the individual and reward-region variability that the neuroforecasting effect depends on (§2) . We run TRI

FAQs

How does the paper test whether TRIBE-predicted neural signals forecast YouTube re-watching behavior?

TRIBE predictions are reduced to a per-second engagement curve (global field power) and correlated with each video’s “most replayed” heatmap, a proxy for moments viewers return to.

What statistical evidence supports the main finding of a null relationship?

The reported pooled partial correlation is near zero and not significantly above loudness or motion baselines; an equivalence test rules out effects larger than about r≈0.14 and a Bayes factor provides moderate evidence for the null.

Is the model’s predicted signal entirely unrelated to re-watch behavior across all analyses?

Most probes show no tracking of re-watch, including across multiple cortical-network readouts, signed value/salience ROIs, and a position-controlled permutation test. A supervised leave-one-video-out probe can show a moderate effect (r≈0.47), but it collapses to a temporal-shape artifact under stronger position control.

Jul 16, 2026

Upload by Ava Thompson