Behavioral variability and neural cross-frequency coupling distinguish working memory maintenance from manipulation

Yingzhe Li, Romy Lorenz, Si WuImaging 2026 · DOI

TL;DR

This study introduces novel, structurally matched tasks to differentiate between working memory (WM) maintenance and manipulation. They found that while both components are related, manipulation involves more complex neural processing, evidenced by distinct brain network activity and a stronger link between neural signals and behavioral performance.

What Problem Does This Paper Address?

Working memory (WM) is a cornerstone of cognition, but its precise structure and how its components—maintenance (holding information) and manipulation (actively processing it)—differ at both behavioral and neural levels remain unclear. Many existing WM tasks are "composite," meaning they tap into both components and potentially other cognitive processes, making it hard to isolate their unique contributions. This lack of clear differentiation hinders our understanding of WM's role in higher-level cognition and the development of targeted interventions. The authors aimed to create a more precise way to measure these components and investigate their distinct neural underpinnings.

Methods

The researchers developed two structurally identical tasks: one for WM maintenance and another for WM manipulation. These tasks used abstract visual stimuli to minimize domain-specific knowledge or strategy use. The key difference was that the manipulation task required participants to reorder the presented visual sequence, adding a layer of active processing beyond simple retention.

To capture neural dynamics, they employed electroencephalography (EEG) and focused on phase-amplitude cross-frequency coupling (CFC). CFC measures how the phase of a slower brain oscillation influences the amplitude of a faster one, thought to be crucial for coordinating neural activity. They analyzed both within-electrode CFC (how a region's own oscillations interact) and cross-electrode CFC (how different brain regions communicate via these coupled oscillations).

Behavioral data (reaction time and accuracy) and neural data (CFC measures) were analyzed using Confirmatory Factor Analysis (CFA) and Structural Equation Modeling (SEM) to assess individual differences and relationships between tasks and neural signals.

Datasets

The study used data collected from 68 participants (aged 18-30) who completed the newly designed WM tasks and a standard operation span task. No external datasets were used for training or evaluation.

Key Insights and Ideas

  • WM Maintenance and Manipulation are Distinct but Related: The study provides strong evidence that while maintenance and manipulation are related, they represent separable cognitive constructs with distinct behavioral and neural signatures.
  • Manipulation is More Demanding: The manipulation task consistently required longer reaction times and engaged more complex neural networks, as indicated by stronger and more widespread CFC.
  • Neural Signals Predict Behavior Differentially: Neural activity (CFC) was a more reliable predictor of performance in the manipulation task than in the maintenance task, suggesting that these neural dynamics are more critical for active processing.
  • CFC as a Window into Neural Dynamics: Phase-amplitude CFC proved to be a sensitive measure for distinguishing between the two WM components, revealing differences in both the strength and the network patterns of neural coordination.

Key Results

  • Behavioral Distinction: Participants were significantly slower in the manipulation task (mean difference of 1.27 seconds) compared to the maintenance task, with high accuracy in both.
  • Neural Distinction: WM manipulation showed significantly higher phase-amplitude CFC (measured by Modulation Index, MI) compared to maintenance.
  • Network Differences: Manipulation recruited more extensive and structurally distinct cross-electrode CFC networks, indicating a more complex pattern of neural communication.
  • Predictive Power: Individual differences in manipulation task performance were reliably predicted by CFC patterns, whereas maintenance performance was not.
  • Factor Analysis: CFA and EFA supported a two-factor model, distinguishing maintenance and manipulation as separate underlying constructs.

Limitations and Future Work

  • Visual Modality Focus: The study primarily used visual stimuli, and future research should explore other sensory modalities to assess the generalizability of these findings.
  • Task Design Nuances: The authors acknowledge that the manipulation task, while unified, might still allow for some inference strategies (e.g., memorizing only two out of three images). Future versions could use four images to mitigate this.
  • Operation Span Task Complexity: The operation span task, while a standard measure, is highly composite and may introduce variability not directly related to core WM processes, leading to weaker correlations with the newly designed tasks.
  • EEG Limitations: The spatial resolution of EEG limits precise inference of neural sources. Future studies could combine EEG with fMRI or intracranial EEG for better anatomical localization.
  • CFC Measurement: While robust, the use of gamma-band activity in EEG can be susceptible to noise.
  • Future Directions: The authors suggest refining WM tasks, exploring cross-sensory modalities, and integrating neural measures for more comprehensive understanding. They also highlight CFC as a promising tool for assessing WM and developing targeted training paradigms.

Summary updated

Abstract

Working memory (WM) is typically understood to comprise two core components: maintenance (short-term retention) and manipulation (active processing of information). However, many existing WM tasks are composite in design, involving both components and even other cognitive processes or constructs. Explicit characterization of the two components in terms of cross-individual variability in both behavioral performance and neural correlates remains limited. To address this gap, we developed two structurally matched task paradigms to compare the behavioral variability in WM maintenance and manipulation while minimizing extraneous influences such as strategy use and domain-specific skills. We also used phase-amplitude cross-frequency coupling (CFC) as a neural measure to investigate the neural dynamics of each WM component. Our psychometric analyses converged in supporting the differentiation between the two components. In addition, neural evidence showed that they engaged different brain networks. Specifically, WM manipulation recruited more extensive networks, reflected in stronger and structurally distinct phase-amplitude CFC across brain regions. Moreover, individual differences in manipulation task performance were reliably predicted by the corresponding CFC patterns, whereas maintenance task performance was not, further highlighting the two components' differential association with neural dynamics.
BibTeX
@inproceedings{li2026behavioral,
  author = {Yingzhe Li and Romy Lorenz and Si Wu},
  title = {{Behavioral variability and neural cross-frequency coupling distinguish working memory maintenance from manipulation}},
  booktitle = {Imaging},
  year = {2026},
  doi = {10.1162/imag.a.1155},
}
Figure