Beyond the control bottleneck: cognitive control reconfigures its neural and computational architecture to meet task demands
Abstract
Cognitive control enables flexible adaptation of behavior, but it has associated costs: task rules must be held in memory and updated when they change. It remains unclear whether these costs arise from a unitary control bottleneck or from the interaction of multiple computational processes, and how such processes are implemented in the brain. Here we recast rule maintenance and switching within a computational model based on probabilistic (active) inference, which we apply to trial-by-trial behavioral and EEG data from a cued task-switching paradigm. Behavior showed a speed–accuracy dissociation, with slower responses during rule repetition and reduced accuracy during rule switching. We identified three separable model parameters as necessary to explain behavior: context stability, cue sensitivity and action precision. These parameters were dynamically reweighted across task contexts, with context stability predominating during rule maintenance and cue sensitivity during rule updating. They also influenced behavior through dissociable routes: higher action precision selectively improved accuracy without slowing responses. Using hierarchical Parametric Empirical Bayes with out-of-sample cross-validation, we related these latent parameters to electrophysiological signals, revealing both context-invariant and context-specific mappings: parietal P3b latency tracked cue sensitivity across conditions, whereas frontal beta latency and the frontal P3a indexed context stability in a context-dependent way. The neural implementation of context stability reorganized with task demands, shifting from frontal beta during switching to frontal theta during maintenance. Together, these findings move beyond the bottleneck view, suggesting that cognitive control is best understood as an adaptive computational architecture that integrates endogenous and exogenous information while reconfiguring its neural implementation according to task demands.
Related
- Mnemonic Demand Reconfigures the Neural Architecture of Memory Encoding
- Emerging Insights into Cognitive Control from Computational Models of Choice
- Reusable modular architecture enables flexible cognitive operations in the mouse brain and artificial recurrent networks
- A hierarchical memory architecture overcomes context limits in long-horizon multi-agent computational modeling
- Ten simple rules for achieving computational reproducibility in cognitive neuroscience
BibTeX
@misc{proietti2026beyond,
title = {{Beyond the control bottleneck: cognitive control reconfigures its neural and computational architecture to meet task demands}},
author = {Riccardo Proietti and Francisco Barcelo and Giovanni Pezzulo},
howpublished = {PsyArXiv},
year = {2026},
month = sep,
doi = {10.31234/osf.io/ugnr6_v1},
url = {https://osf.io/preprints/psyarxiv/ugnr6_v1/},
note = {CC0},
}
RIS
TY - UNPB
TI - Beyond the control bottleneck: cognitive control reconfigures its neural and computational architecture to meet task demands
AU - Riccardo Proietti
AU - Francisco Barcelo
AU - Giovanni Pezzulo
PY - 2026
DA - 2026/09/10
PB - PsyArXiv
DO - 10.31234/osf.io/ugnr6_v1
UR - https://osf.io/preprints/psyarxiv/ugnr6_v1/
AB - Cognitive control enables flexible adaptation of behavior, but it has associated costs: task rules must be held in memory and updated when they change. It remains unclear whether these costs arise from a unitary control bottleneck or from the interaction of multiple computational processes, and how such processes are implemented in the brain. Here we recast rule maintenance and switching within a computational model based on probabilistic (active) inference, which we apply to trial-by-trial behavioral and EEG data from a cued task-switching paradigm. Behavior showed a speed–accuracy dissociation, with slower responses during rule repetition and reduced accuracy during rule switching. We identified three separable model parameters as necessary to explain behavior: context stability, cue sensitivity and action precision. These parameters were dynamically reweighted across task contexts, with context stability predominating during rule maintenance and cue sensitivity during rule updating. They also influenced behavior through dissociable routes: higher action precision selectively improved accuracy without slowing responses. Using hierarchical Parametric Empirical Bayes with out-of-sample cross-validation, we related these latent parameters to electrophysiological signals, revealing both context-invariant and context-specific mappings: parietal P3b latency tracked cue sensitivity across conditions, whereas frontal beta latency and the frontal P3a indexed context stability in a context-dependent way. The neural implementation of context stability reorganized with task demands, shifting from frontal beta during switching to frontal theta during maintenance. Together, these findings move beyond the bottleneck view, suggesting that cognitive control is best understood as an adaptive computational architecture that integrates endogenous and exogenous information while reconfiguring its neural implementation according to task demands.
ER -