Sequential Contextual Fit Predicts Human Behavioural and Neural Dynamics Across Domains
cs.AI · cs.LG
Quick verdict
Recency-weighted embedding-similarity operator predicts reading, emotion, choice and neural dynamics beyond domain baselines; broad evidence, modest novelty.
Summary
The paper introduces sequential contextual fit (SCF), a measurement operator computed as the recency-weighted average cosine similarity between a unit's embedding and the embeddings of the previous K units. The authors apply this single operator to public datasets from six domains: MECO eye-tracking (13 languages), DERCo reading EEG, Alice/Moth fMRI, DEAM music emotion, EAV audiovisual emotion EEG (participants 1-10 only), Many Labs Iowa Gambling Task, UCI HAPT activity data, and a decision EEG dataset. Across domains, lower SCF predicts longer fixation durations beyond surprisal (GAMM F up to 26.8), broader FDR-corrected ERP coverage than surprisal in N400/P600 windows, fMRI SCF effects in all 12 Alice ROIs where surprisal was null, larger next-step affective change beyond MERT acoustic change (F=89.13, dAIC=294.52), next-trial switching beyond negative prediction error (chi-squared=654.39, dAIC=580.25), postural transitions beyond sensor change (chi-squared=300.31, dAIC=458.79), and EEG-state transitions (F=5.83). Robustness checks include kernel variants, K in {1,3,4,5}, grouped leave-subject-out CV, and AR(1) checks. Key limitations: the operator closely resembles prior self-similarity/novelty measures and the authors' own published semantic-relevance work; the IGT state embedding contains choice-related features, risking near-tautological switch prediction; held-out emotion prediction gains are near zero (R-squared roughly 0.0005); and the EAV subset selection is unexplained.
Review complete
Assessment at a glance
6
risk flags identified
● high 2
● medium 3
● low 1
5 strengths · 5 weaknesses · 5 author questions
AI text signal = Binoculars zero-shot detector score over 20 sampled passages (Fast-DetectGPT co-reported) — absence does not prove human authorship, and a flag does not imply misconduct.
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- SCF, a recency-weighted average cosine similarity between a unit's embedding and the previous K units' embeddings, is applicable across words, audio, video frames, choices, sensor windows and neural states.
- Lower SCF predicts longer fixation durations across 13 languages beyond transformer surprisal, word frequency and length (MECO GAMMs).
- SCF explains word-locked EEG (N400/P600) and fMRI BOLD responses beyond surprisal, with surprisal null in all 12 Alice ROIs under FIR/deconvolution.
- Lower SCF predicts larger next-step affective change (DEAM), next-trial switching (IGT), postural transitions (HAPT) and single-trial EEG-state transitions beyond domain baselines.
- Effects are robust across context windows K=1-5 and kernel variants and yield leave-subject-out prediction gains in action and decision domains.
Strengths
Genuine cross-domain breadth on public data
One operator applied to nine analyses over MECO (13 languages), DERCo, Alice/Moth, DEAM, EAV, Many Labs IGT, UCI HAPT and decision EEG (Table 1), with a consistent low-fit-to-more-updating direction across symbolic, affective, choice and sensor domains.
Robustness battery above field norm
Kernel variants (unweighted, pairwise, centroid; Sec. 2.3), window sensitivity K=1,3,4,5 with all multi-unit windows significant (Sec. 3.5), grouped five-fold leave-subject-out CV, and AR(1)/block checks for DEAM and HAPT reported in the appendix.
Anti-circularity design in EAV
Baseline-contested joint models
Transparent, parameter-free metric definition
Weaknesses
Thin novelty with heavy self-overlap
Near-tautological IGT result
In-sample significance vs near-zero generalization in emotion
Unexplained EAV subset
Weak decision baseline and heterogeneous operator
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Scope & Claims
The paper defines sequential contextual fit (SCF) as the recency-weighted average cosine similarity between a current unit's embedding and the embeddings of the previous K units (Eqs. 1-3; linear kernel a_i=(K-i+1)/K), and positions it as a domain-general measurement layer rather than a generative theory. Seven analysis families span six domains (Table 1): MECO eye-tracking in 13 languages, DERCo reading EEG, Alice and Moth naturalistic-listening fMRI, DEAM music emotion, EAV audiovisual emotion EEG (participants 1-10 of 42), Many Labs Iowa Gambling Task behaviour, UCI HAPT activity recognition, and a separate IGT EEG dataset. The central claim is that lower SCF predicts longer processing times, larger affective/behavioural transitions and stronger neural-state changes beyond domain baselines (surprisal, negative RL prediction error, acoustic/visual/sensor change).
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 5 strengths & 5 weaknesses, analyzed in detail
- 6 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
What is sequential contextual fit (SCF)?
SCF is a simple metric that measures how well the current element of a sequence (a word, sound segment, video frame, choice, sensor window or brain state, represented as a vector embedding) matches its recent context. It is computed as the average cosine similarity between the current vector and the previous K vectors, weighted so that more recent items count more (a linear recency kernel). High SCF means representational continuity with recent context; low SCF means mismatch. It requires no fitted probability, reward or hidden-state model.
Is SCF better than surprisal for explaining reading behavior?
The paper argues SCF is complementary, not superior. In joint GAMM models on the MECO eye-tracking corpus (13 languages), SCF predicted first fixation, gaze and total fixation durations (F values up to about 26.8) even after controlling for transformer-based surprisal, word frequency and length, and showed broader FDR-corrected EEG coverage than surprisal in N400/P600 windows. In the Alice fMRI data, SCF was significant in all 12 ROIs while surprisal was not significant in any under their FIR analysis — a striking result that would need independent verification given prior literature showing surprisal effects.
What data does the paper use and is code available?
All analyses use public datasets: MECO (eye-tracking), DERCo (reading EEG), Alice and Moth (naturalistic-listening fMRI), DEAM (music emotion ratings), EAV (audiovisual emotion EEG, but only participants 1-10 of 42), the Many Labs Iowa Gambling Task behavioural data, the UCI HAPT smartphone-sensor corpus and a public IGT EEG dataset. A Reproducibility section states that a project archive contains processed datasets, embedding/SCF construction scripts, GAMM scripts and saved model outputs (Python 3.10.11, R 4.5.1, mgcv 1.9.4); the actual archive link is deferred to the Supplementary Material.
Does SCF actually predict anything out of sample?
Partially. In grouped leave-subject-out cross-validation, adding SCF improved held-out AUC for action transitions (0.9743 to about 0.9864 at K=5) and for trial switching (about 0.647 to 0.736-0.739). For music emotion, however, held-out R-squared was essentially zero both with and without SCF (from about -0.002 to at most 0.0005), despite highly significant in-sample F statistics — so out-of-sample value is domain-dependent and weakest exactly where the in-sample evidence looks strongest.
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- The extracted text omits several sections (parts of Methods 2.7, portions of the Discussion, details of reference 54); this review reflects the visible text only.
- Figures 1-6 could not be analyzed; statistics were taken from body text and figure captions.
- The Supplementary Material (alternative-metric results, concurvity diagnostics, code archive link) was not available, so several claims are unverifiable from this text.
- Minor metadata inconsistency: arXiv stamp reads 24 Jul 2026 while the manuscript date line reads September 18, 2026.
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.