The paper introduces Sequential Contextual Fit (SCF), an embedding‑based metric that quantifies how well the current information state aligns with its recent context. SCF is computed with a simple recency‑weighted similarity kernel: $$\text{SCF}_t = \sum_{\tau=1}^{T}\alpha_{\tau}\,\text{sim}(\mathbf{x}_t, \mathbf{x}_{t-\tau})$@@@MATH_BLOCK1@@@\alpha{\tau}=\exp(-\lambda\tau)$ provides exponential decay over past time steps and $\text{sim}$ is cosine similarity.
SCF can be applied uniformly to words, sounds, visual scenes, affective states, choices, actions, and neural representations, without domain‑specific tuning. The authors evaluated SCF across six datasets: language processing, music‑evoked emotion, a subset of audiovisual emotion EEG, gambling decisions, human activity recognition, and decision‑related EEG. Across all domains, lower contextual fit consistently predicted longer processing times, larger affective or behavioural transitions, and stronger neural‑state changes.
Crucially, each analysis controlled for established predictors such as surprisal, reinforcement‑learning prediction error, acoustic change, visual change, and sensor change. Even after accounting for these factors, SCF remained a significant predictor of residual behavioural and neural variance.
Thus, SCF offers a universal computational layer that links contextual compatibility to behavioral latency, cognitive/neuronal state transitions, and can serve as a common analytical tool for cross‑domain cognitive neuroscience research.
Review