Neural Mechanisms of Willed Attention Control

Article activity feed

1. 

   eLife

   Nov 10, 2026

   eLife Assessment

   This important study advances our understanding of how ongoing brain activity relates to people's choices about where to direct attention. While the evidence from the two independent datasets combining fMRI and EEG under the same willed-attention paradigm is solid, significant concerns remain regarding the rigor of the classifier analyses. This work will interest cognitive psychologists and neuroscience researchers.

   Read the original source
2. 

   eLife

   Nov 10, 2026

   Reviewer #1 (Public review):

   Summary:

   This study addresses a fundamental question in cognitive neuroscience regarding how the brain transitions from a reactive state of following external instructions to a proactive state of self-directed agency. The authors utilize an ambitious multimodal design by combining the spatial precision of fMRI with the temporal resolution of EEG across two independent datasets from the University of Florida and UC Davis. The work shows that while both instructed and willed attention engage the Dorsal Attention Network, willed choices additionally recruit a frontoparietal network including the dACC and anterior insula. By applying multivariate pattern analysis, the authors further demonstrate that the chosen direction of attention can be decoded from this network only on choice trials, and that pre-cue alpha …

   Reviewer #1 (Public review):

   Summary:

   This study addresses a fundamental question in cognitive neuroscience regarding how the brain transitions from a reactive state of following external instructions to a proactive state of self-directed agency. The authors utilize an ambitious multimodal design by combining the spatial precision of fMRI with the temporal resolution of EEG across two independent datasets from the University of Florida and UC Davis. The work shows that while both instructed and willed attention engage the Dorsal Attention Network, willed choices additionally recruit a frontoparietal network including the dACC and anterior insula. By applying multivariate pattern analysis, the authors further demonstrate that the chosen direction of attention can be decoded from this network only on choice trials, and that pre-cue alpha patterns carry information predictive of subsequent spontaneous choices. This suggests a neural link between pre-existing brain states and intentional action, representing a significant technical effort to characterize the neural scaffolding of internal goal generation.

   Strengths:

   The primary strengths of this work include the integration of fMRI and EEG, which allows the authors to bridge the gap between slow metabolic signals and fast oscillatory brain states. The use of two independent cohorts is a commendable effort to ensure the reproducibility of the willed attention effect, which is often a concern in small-sample neuroimaging studies. Additionally, the move beyond univariate activation toward information-based mapping remains the strongest aspect of the work, as it shows that the identified network actually contains specific information about the direction of attention. However, the per-site decoding effects are modest and, in the UF dataset alone, do not survive correction (Table A1). The revision further adds permutation- and cluster-based inference, channel-subset and trial-balanced EEG decoding, and clearer methodological reporting, all of which improve transparency.

   Weaknesses:

   The revised manuscript addresses most of my earlier concerns, and the additional analyses improve the transparency of the work. The remaining points concern interpretation rather than the findings themselves. The decoding shows that the frontoparietal network carries the self-chosen direction of attention, but the present data do not establish the temporal ordering proposed in Figure 6. Similarly, the pre-cue alpha effect is consistent with a spontaneous brain-state bias, but without a measure of eye position or of when decisions are made, an oculomotor contribution or choices settled before the cue cannot be excluded. Finally, the link between pre-cue predictability and post-cue frontoparietal activity rests on a between-subject comparison, and the neural efficiency interpretation is one of several possible readings. Readers should weigh the proposed model with these limitations in mind.

   Read the original source
3. 

   eLife

   Nov 10, 2026

   Reviewer #2 (Public review):

   Summary:

   This manuscript combines fMRI and EEG investigations performed at two research sites to examine 'willed' or volitional visuospatial attention, as contrasted with more standard cued (or 'instructed') visuospatial attention. The primary findings are: 1) willed attention (vs. instructed attention) drives additional cortical circuitry across a broad fronto-parietal network; 2) the direction of willed attention, but not instructed attention can be decoded from the pre-cue EEG data and from MVPA analysis of the trial-level fMRI data; and 3) the subjects with high EEG decoding also exhibited high neural efficiency (i.e., high decoding with low BOLD signal change) in the fMRI data. There are significant concerns regarding the statistical rigor of the direction decoding analyses. These were incompletely …

   Reviewer #2 (Public review):

   Summary:

   This manuscript combines fMRI and EEG investigations performed at two research sites to examine 'willed' or volitional visuospatial attention, as contrasted with more standard cued (or 'instructed') visuospatial attention. The primary findings are: 1) willed attention (vs. instructed attention) drives additional cortical circuitry across a broad fronto-parietal network; 2) the direction of willed attention, but not instructed attention can be decoded from the pre-cue EEG data and from MVPA analysis of the trial-level fMRI data; and 3) the subjects with high EEG decoding also exhibited high neural efficiency (i.e., high decoding with low BOLD signal change) in the fMRI data. There are significant concerns regarding the statistical rigor of the direction decoding analyses. These were incompletely addressed in the revision. These concerns are laid out in more detail below. Additionally, there is a concern over the degree to which the present research paradigm adequately examines 'willed attention,' which may mute the potential impact.

   Strengths:

   The focus on willed attention attempts to move beyond some of the many limitations of standard laboratory investigations of attention.

   The shared paradigm across two modalities and two research sites demonstrates solid reproducibility, even though a few minor differences are observed across sites.

   Weaknesses:

   Major Concerns:

   This revised manuscript addresses most of the concerns raised in the first set of reviews, however significant concerns still remain with regards to the 'multivariate analysis of cue-evoked BOLD activation' methods and results. This is central to many of the novel claims of this manuscript, therefore it is hoped that the authors will address this further. Two key concerns were raised by reviewers 2 and 3 about the decoding analyses reported in Figures 3 and 4. Although these concerns were addressed separately, both concerns impact both the fMRI and alpha power analyses but were not addressed within a single analysis for each data set. This appears to leave the critical decoding analysis results unresolved. The revised manuscript still states "For both datasets, decoding accuracy for the instructed trials was not significantly different from chance level (50%) for all the ROIs." However, the 50% level is not the appropriate comparison for statistical testing. Specifically, Figures 3 and 4 should be reproduced with analyses that include the appropriate corrections and controls.

   There are 2 key issues: 1) Many individual subjects did not choose left or right at a 50-50 probability and these subject-level response biases provide a trivial means by which a classifier can perform above 50%; 2) Classifiers can 'perform above chance, by chance.' Empirical estimates of 'chance', generated by label swapping permutations are strongly recommended. Both issues are critical for both the MVPA and alpha power decoding analysis. In their supplemental materials of the revision the authors address point 1 for the alpha power decoding and point 2 for the MVPA decoding, but they do not address both issues together for either data set. This needs to be done if we are to believe the decoding results. Each concern appears to reduce classifier performance or significance; therefore, it is not clear which results would survive joint application of correction methods.

   Figure 4: This reviewer (R2) raised concerns that individual subjects likely exhibited unequal numbers of right and left decisions, which would allow a simple classifier to perform > 50% by always guessing the biased direction. Figure A3a demonstrates that indeed many subjects exhibited biases in one direction or the other. The fact that these biases average out to close to 50% across the subject pool (Figure 3a) is irrelevant to the decoder analysis, which is performed at the individual subject level. The analysis presented for the alpha power decoding in Figure A3b addresses the imbalance by throwing out trials to balance the right/left decisions within subjects. This seems like a reasonable approach. The net effect is that the overall classifier performance drops from 56.83% accuracy to 52.81%, confirming that subject response biases were a large factor in MVPA performance. In support of the author claims, 52.81% performance is found to be significantly greater than chance in the combined data set (p=0.028). The results of this analysis should replace what is currently displayed in Figure 4, and also the contents of Figure A1. Since Figure 4 divides the data into site-specific results and into time courses, it is far from clear which of these results will remain significant.

   Figure 3: The subject-level response biases are not addressed within the fMRI decoding analysis of Figure 3 but should be. Then the permutation analysis of Table A1 should be performed. Or an alternate approach could be taken that addresses both issues simultaneously. The results of the combined analysis should replace what is reported in Figure 3, which very likely overstates the effects.

   Discussion: There are concerns about this experimental paradigm carrying the banner of Willed Attention, because the application of 'Will' appears quite modest. Yes, extra brain activity is exhibited for this condition vs. its control, but do the cognitive processes isolated adequately stand in for 'Willed Attention?" Willed attention, as operationally defined here, appears to involve a simple decision process prior to the shifting of spatial attention. In this revision, the authors have provided additional context that should aid future studies.

   Read the original source
4. 

   eLife

   Nov 10, 2026

   Reviewer #3 (Public review):

   Summary:

   This study uses two independent neuroimaging datasets to investigate the neural correlates of internally chosen spatial attention. The authors examine choice-related information in the frontoparietal regions and prediction of subsequent choices from the precue alpha patterns. They also examined the associations between precue predictability and postcue network responses. The revision and supplementary analyses improve methodological transparency and strengthen the evidence for precue alpha prediction.

   Strengths:

   The question is worthwhile, and the two independent datasets provide a useful assessment of consistency across sites. EEG and fMRI offer complementary information about precue activity and postcue spatial representations.

   The revision clarifies ROI definitions and statistical testing, …

   Reviewer #3 (Public review):

   Summary:

   This study uses two independent neuroimaging datasets to investigate the neural correlates of internally chosen spatial attention. The authors examine choice-related information in the frontoparietal regions and prediction of subsequent choices from the precue alpha patterns. They also examined the associations between precue predictability and postcue network responses. The revision and supplementary analyses improve methodological transparency and strengthen the evidence for precue alpha prediction.

   Strengths:

   The question is worthwhile, and the two independent datasets provide a useful assessment of consistency across sites. EEG and fMRI offer complementary information about precue activity and postcue spatial representations.

   The revision clarifies ROI definitions and statistical testing, provides permutation-based fMRI results and effect sizes, and introduces cluster-based correction for the EEG time course. Importantly, precue alpha decoding remains above the empirical permutation baseline after balancing left- and right-choice trials within participants. This directly addresses the concern that individual choice biases could account for above-chance accuracy. Analyses restricted to posterior or frontal electrodes also support the conclusion that the predictive effect does not depend on using the full electrode set.

   Read the original source
5. 

   eLife

   Nov 10, 2026

   Author response:

   The following is the authors’ response to the original reviews.

   Public Reviews:

   Reviewer #1 (Public review):

   Summary:

   This study addresses a fundamental question in cognitive neuroscience regarding how the brain transitions from a reactive state of following external instructions to a proactive state of self-directed agency. The authors utilize an ambitious multimodal design by combining the spatial precision of fMRI with the temporal resolution of EEG across two independent datasets from the University of Florida and UC Davis. By applying multivariate pattern analysis, the work demonstrates that while both instructed and willed attention engage the Dorsal Attention Network, willed choices uniquely recruit a frontoparietal decision network including the dACC and anterior insula. Furthermore, the study shows that …

   Author response:

   The following is the authors’ response to the original reviews.

   Public Reviews:

   Reviewer #1 (Public review):

   Summary:

   This study addresses a fundamental question in cognitive neuroscience regarding how the brain transitions from a reactive state of following external instructions to a proactive state of self-directed agency. The authors utilize an ambitious multimodal design by combining the spatial precision of fMRI with the temporal resolution of EEG across two independent datasets from the University of Florida and UC Davis. By applying multivariate pattern analysis, the work demonstrates that while both instructed and willed attention engage the Dorsal Attention Network, willed choices uniquely recruit a frontoparietal decision network including the dACC and anterior insula. Furthermore, the study shows that pre-cue alpha oscillations can predict subsequent spontaneous choices. This provides a neural link between pre-existing brain states and intentional action, representing a significant technical effort to characterize the neural scaffolding of internal goal generation.

   Thank you for the succinct summary of our work.

   Strengths:

   The primary strengths of this work include the integration of fMRI and EEG which allows the authors to bridge the gap between slow metabolic signals and fast oscillatory brain states. The use of two independent cohorts is a commendable effort to ensure the reproducibility of the willed attention effect, which is often a concern in small-sample neuroimaging studies. Additionally, the move beyond univariate activation toward information-based mapping demonstrates that the identified networks actually contain specific information about the direction of attention.

   Thank you for the positive assessment of our work.

   Weaknesses:

   (1) However, several critical weaknesses must be addressed to support the fundamental claims made in the manuscript. There are significant behavioral differences in performance between the two sites, specifically regarding the UC Davis cohort exhibiting slower reaction times and lower accuracy compared to the UF group. These discrepancies suggest potential differences in subject populations or experimental environments that are not currently accounted for in the neural models. The fMRI analysis lacks temporal precision because the use of beta series regression collapses the complex BOLD response into a single estimate per trial. This loss of temporal information obscures the evolution of the decision process and makes it difficult to distinguish whether the identified patterns represent a truly spontaneous choice or a slow-building, pre-planned strategy.

   The reviewer is correct to point out the quantitative differences in behavioral performance between the UF and UCD cohorts, with the UCD participants exhibiting slower reaction times and somewhat lower accuracy despite the use of the same willed‑attention paradigm and matched stimulus parameters. The exact source of these differences is not fully known, but one likely contributor is that at UF, EEG and fMRI were recorded simultaneously in a single session, whereas at UCD, EEG and fMRI were acquired in separate sessions, which may have affected overall task engagement and fatigue. To address this concern in our analyses, we (i) first examined each dataset separately, showing that the main qualitative findings (choice‑cue activation of the frontoparietal decision network, decoding of choice in frontoparietal regions, and precue alpha predictability of willed choices) replicate across sites, and (ii) then combined datasets only at a higher level (using meta‑analytic/summary‑statistic approaches) to increase statistical power while respecting site‑specific differences. We note the behavioral differences and emphasize the analysis strategy in the revised manuscript.

   Regarding the temporal precision of the fMRI analysis, we agree that single‑trial beta estimates do not allow us to track the evolution of the decision process on a sub‑second timescale and thus cannot, by themselves, distinguish a truly spontaneous choice from a slowly developing pre‑planned strategy. Our fMRI analyses were designed to identify where willed versus instructed attention engage distinct control networks and encode decision‑relevant information, rather than to resolve the precise timing of decision formation. The temporal aspects of the decision process in this paradigm have been addressed in prior work using EEG (Rajan et al., 2018). In that study, we found that frontal theta oscillations serve as an index of decision‑making: for choice trials, a burst of frontal theta power emerged between approximately 400 and 800 ms after the cue, a pattern that was not present in instructed trials. In addition, using the onset of posterior alpha lateralization as a marker of attentional set implementation, we showed that in choice trials the deployment of the posterior attention set was delayed by about 400–500 ms relative to instructed trials. Together, these EEG findings suggest that the decision process takes place in the time period of 400 to 800 ms after the choice cue and the duration of the process is in the neighborhood of 400–500 ms, supporting the view that the decision about where to attend are formed during the cue–target interval, rather than being fully pre‑planned well in advance. We now discuss the temporal dynamics more explicitly in the revised manuscript.

   (2) Furthermore, the EEG decoding approach utilized the entire topography of electrodes rather than a biologically motivated posterior region of interest. Given that alpha-mediated spatial attention is traditionally localized to parieto-occipital sensors, using the full electrode set risks the inclusion of non-neural artefacts such as microsaccades or muscle activity which can contaminate multivariate classifiers. The introduction of the neural efficiency metric also requires further validation as the current ratio is mathematically sensitive to small denominators in the BOLD contrast.

   The reviewer’s point about the decoding topography is well taken. To assess whether our results depend on using the full electrode set, we re-ran the precue alpha decoding using only posterior parieto‑occipital electrodes (posterior ROI) and, separately, using only frontal electrodes (frontal ROI). In Figure A1, all three analyses (whole‑head, posterior‑only, and frontal‑only) yielded a similar pattern: for choice trials, decoding accuracies were reliably above chance, whereas for instructed trials, decoding accuracies remained at chance. Thus, the key qualitative effect does not depend on including non‑posterior channels.

   We chose the whole‑topography approach because our question concerned distributed alpha patterns rather than strictly local effects. In particular, frontal structures such as dACC and DLPFC are known to monitor internal brain states and interact with posterior attention networks, making it plausible that both posterior and frontal channels can contribute to decoding the upcoming decision about where to attend. At the same time, we agree that artefacts are a concern in multivariate EEG analyses. In the present case, the decoding window is strictly precue. Given that instructed and choice trials are interleaved and subjected to identical preprocessing and artefact rejection procedures, it is unlikely that systematic differences in eye or muscle artefacts drive the choice‑specific decoding. We now report the posterior‑only and frontal‑only decoding results in the Supplementary Materials and comment on the result in the revised manuscript.

   Regarding the neural efficiency metric, we agree that ratio measures can, in principle, be sensitive to small denominators. In the present data, the BOLD contrasts used in the denominator are significantly different from zero at the group level, and we interpret the neural efficiency metric as a descriptive summary of the relationship between decoding and BOLD responses rather than as a standalone biomarker. We now explicitly note this potential sensitivity and limitation in the revised manuscript.

   (3) Crucially, the manuscript does not address the physiological implications of recruiting additional frontoparietal networks when behavioral performance remains identical across conditions. The activation of the anterior insula and dACC is frequently associated with increased autonomic arousal and effort. If the willed condition requires more extensive neural scaffolding to reach the same behavioral output as the instructed condition, it raises the question of whether this internal decision process is accompanied by changes in arousal levels. The authors should consider whether the lack of a behavioral tax is due to a compensatory increase in arousal, which could be reflected in the EEG data or pupil diameter if recorded, and potentially also in the amplitude of BOLD activity, which is being masked by the neural efficiency metric. Without an account of how the brain balances this increased computational demand without impacting behavioral performance, the functional significance of the willed attention network remains partially obscured.

   We thank the reviewer for this comment, which prompted us to examine the behavioral and physiological data more closely for evidence of a compensatory arousal or effort mechanism.

   Regarding behavior, Table 1 shows that choice trials are associated with longer reaction times but higher accuracy relative to matched instructed trials (choose-left vs. cue-left, choose-right vs. cue-right). These effects were not consistently significant across both measures and both sites (UF: RT p = .438, accuracy p = .012; UCD: RT p = .830, accuracy p = .292), so we interpret this pattern cautiously rather than as a fully established effect. Nonetheless, the direction of the pattern—slower but more accurate responses on choice trials—is consistent with a speed–accuracy trade-off associated with increased cognitive control, rather than with a straightforward behavioral cost of engaging the willed attention network. This supports the interpretation that additional recruitment of anterior insula and dACC may reflect increased effortful control that helps offset, rather than degrade, task performance.

   Regarding physiological arousal more directly, pupillometry was not recorded in this study, so we cannot speak to pupil-based arousal measures. However, following the reviewer's suggestion, we examined the time course of occipital alpha power (averaged across O1, O2, and Oz, using 500-ms sliding windows in 20-ms steps, expressed as dB change from pre-cue baseline) as an EEG-based index of attentional/arousal-related engagement. In Figure A2, alpha suppression 600–1200 ms after cue onset was significantly stronger for choice trials than instructed trials in the UF dataset (p = .030), with a difference in the same direction, though not statistically significant, in the UCD dataset (p = .163). Stronger alpha suppression is generally taken to reflect greater cortical engagement/arousal, so this pattern is broadly consistent with the idea that willed attention trials involve a heightened engagement state relative to instructed trials.

   Together, these behavioral and EEG results provide converging, though not uniformly robust, evidence that the additional frontoparietal (and salience network) recruitment associated with willed attention may be accompanied by a compensatory increase in effort/arousal-related engagement, which could help explain why behavioral performance is not impaired—and if anything trends toward improved accuracy—in the choice condition. We now include Figure A2 in the Supplementary Materials and note that direct physiological arousal measures (e.g., pupillometry) were not available in the present dataset and would be a valuable addition in future work directly designed to test this hypothesis.

   Reviewer #2 (Public review):

   Summary:

   This manuscript combines fMRI and EEG investigations performed at two research sites to examine 'willed' or volitional visuospatial attention, as contrasted with more standard cued (or 'instructed') visuospatial attention. The primary findings are: 1) willed attention (vs. instructed attention) drives additional cortical circuitry across a broad fronto-parietal network; 2) the direction of willed attention, but not instructed attention can be decoded from the pre-cue EEG data and from MVPA analysis of the trial-level fMRI data; and 3) the subjects with high EEG decoding also exhibited high neural efficiency (i.e., high decoding with low BOLD signal change) in the fMRI data. The methods and data analysis are generally sound, and these results appear solid. On the negative side, it is not made clear how the present findings extend our understanding beyond prior published work from one of the senior authors. There are also three significant concerns regarding interpretation of the findings. One has to do with the causal interpretation of the pre-cue alpha EEG signal determining the direction of willed attention. The second concern is the degree to which the present research paradigm adequately examines 'willed attention.' The third is that the MVPA analysis is not sufficiently described, and Permutation testing needs to be done to validate these findings. Otherwise, this manuscript appears methodologically sound, but questions about interpretation may mute the potential impact.

   Thank you for the succinct summary of our work. We address the concerns below.

   Strengths:

   The focus on willed attention attempts to move beyond some of the many limitations of standard laboratory investigations of attention.

   The shared paradigm across two modalities and two research sites demonstrates solid reproducibility, even though a few minor differences are observed across sites.

   Thank you for the positive assessment of our work.

   Weaknesses:

   (1) There are concerns about this experimental paradigm carrying the banner of Willed Attention, because the application of 'Will' appears quite modest. Yes, extra brain activity is exhibited for this condition vs. its control, but do the cognitive processes isolated adequately stand in for 'Willed Attention?" Willed attention, as operationally defined here, appears to involve a simple decision process prior to the shifting of spatial attention. The cue is internally generated, but after that the rest of the attentional processes appear identical to standard externally cued visuospatial attention experiments. This self-generated cue process likely involves some sort of memory/history of the recently selected cues and then some randomish selection between A and B. This appears very similar to asking the subject to guess whether a fair coin flip will be heads or tails on each trial. A mental 'coin flip' feels like a very weak version of 'will.' As a potential remedy, it would be helpful to discuss what other phenomena might fall within 'willed attention' and what some future studies might choose to focus on, along with some potential pitfalls (e.g., the reasons why the current study avoided more robust exemplars of will).

   Thank you for this thoughtful comment about the implications of “willed attention” as instantiated in our paradigm. Our use of the term follows earlier work in this line of research, where attention that is endogenously directed in the absence of external attention‑directing cues has been referred to as “willed attention” (Bengson et al., 2014, 2015; Liu et al., 2017; Rajan et al., 2018; see also Nadra et al., 2023, for review). In these studies, and in the present work, participants are prompted to decide for themselves where to attend (left vs right) when a choice cue appears, rather than being instructed where to attend by an external cue.

   We agree that, viewed from a broader perspective on “will,” this is a modest instantiation. The choice component can be described as a simple internally generated decision between two alternatives, and we do not claim that it captures the full richness of everyday volitional attention (e.g., spontaneously deciding what to attend, when to shift, or sustaining attention over longer time scales). Our aim here was to isolate, under tightly controlled conditions, the additional control processes that are engaged when the direction of spatial attention is chosen endogenously rather than specified by an external cue. The evidence we present— that the choice cue recruits a frontoparietal decision network, that choice direction can be decoded from this network only in choice trials, and that ongoing alpha patterns bias upcoming choices—speaks directly to this internally generated decision stage within the broader spatial attention framework.

   We have revised the manuscript to emphasize the limited nature of our paradigm and discuss other phenomena that could fall under willed attention, such as self‑initiated shifts of attention without any external prompt, decisions about what feature or object category to attend, or self‑timed initiation and termination of attentional episodes, and moreover, we refer to the review by Nadra and Mangun (2023) where broader issues are discussed.

   (2) The manuscript is lacking a description of the decision processes used during the willed attention paradigm and is lacking evidence as to WHEN subjects made their willed decision. Both of these points are of major concern:

   These are important points. The temporal aspects of the decision process in this paradigm have been addressed in prior work using EEG (Rajan et al., 2018). In that study, we found that frontal theta oscillations serve as an index of decision‑making: for choice trials, a burst of frontal theta power emerged between approximately 400 and 800 ms after the cue, a pattern that was not present in instructed trials. In addition, using the onset of posterior alpha lateralization as a marker of attentional set implementation, we showed that in choice trials, the deployment of the posterior attention set was delayed by about 400–500 ms relative to instructed trials. Together, these EEG findings suggest that the decision process takes place in the time period of ~400–800 ms after the choice cue and it lasts on the order of 400–500 ms. We now explicitly discuss the timing of the decision process in the revised manuscript.

   (a) The authors state: "For willed attention, participants were explicitly told to avoid relying on any stereotypical strategies of generating decisions, such as always attending the same/opposite side they attended during the previous trial, as well as to avoid randomizing or equalizing their decisions to choose left or right across trials; prior studies found that decisions to explicitly randomize decisions might invoke additional working memory related processes (Spence & Frith, 1999)." Subjects were instructed NOT to apply a simple heuristic and NOT to randomize or try to equalize their decisions, but exactly HOW the subjects made their decisions is not at all clear. What options does that leave? How does this strategy avoid the working memory-related processes mentioned in the Spence & Frith, 1999 citation? The brain regions that comprise the network of interest (aka Frontoparietal Decision Network) are activated by a very broad range of visual cognitive tasks, including many working memory paradigms. The Anterior Insula and dACC nodes Salience Network often simply reflect task difficulty. Obviously, making a choice is more cognitively demanding than not making a choice. The present experiments do not distinguish functional roles between different regions of the Frontoparietal Decision Network. On the whole, the study does very little to isolate the cognitive processes or neural bases of willed attention beyond calling out the set of 'Usual Suspects' for visual cognition.

   These are important issues. The purpose of the instructions was to encourage subjects to make a genuinely spontaneous decision about where to attend on each choice trial, rather than relying on simple deterministic strategies (e.g., always alternating left/right) or explicitly trying to randomize or equalize their choices across the session. Spence and Frith (1999) pointed out that attempts to explicitly generate random sequences can recruit working memory and additional control processes. Our instructions were therefore designed to minimize both stereotyped patterns and explicit randomization strategies, and instead to encourage subjects to “go with their momentary choice” on each trial. While we cannot directly observe the exact decision rule each individual used, prior analyses of sequence structure in this paradigm (Bengson et al., 2014) found no evidence that first‑ or second‑order sequence effects (e.g., simple alternation or repetition biases) explained the pattern of choices, consistent with the idea that responses were not governed by a simple fixed strategy.

   Regarding the frontoparietal network, prior work has indeed shown that the choice cue in this task activates regions such as dACC, DLPFC, anterior insula, and IPL (Bengson et al., 2015; Liu et al., 2017). We fully agree with the reviewer that these regions support a wide range of cognitive functions (including working memory, cognitive control, and task difficulty), and we do not claim to have dissociated all of these functions here. The contribution of the present work is to go beyond univariate “activation maps” by showing that: (i) multivoxel patterns in this frontoparietal network encode which spatial alternative is chosen on a given trial (left vs right) in willed attention, whereas no such information is decodable in instructed attention, and (ii) trial‑by‑trial patterns of precue alpha activity predict subsequent choices and frontoparietal activity. Together, these findings link this network not only to general task engagement, but specifically to the internally generated decision about where to attend in this paradigm. We have revised the manuscript to further clarify these points and emphasized that the component nodes of the frontoparietal network are not functionally decomposed in detail by the present study.

   (b) The finding that pre-cue EEG signals predicted the post-cue decision is intriguing. It could mean that the seemingly irrelevant and transient state of the brain causally and unconsciously biased the subject to one direction or the other. Alternatively, it could mean that the subjects utilized the pre-cue period to make their decision and hold it in case it was needed (i.e., that it was a choice trial). While 2-8 seconds ITI variability makes sense for fMRI decoding, it is a long time for a subject to idly wait, so they might fill that time preparing for the next trial. There appears to have been a substantial amount of individual difference in the pre-cue alpha decoding, which could reflect individual differences in cognitive strategy, specifically in the use of the pre-cue period to make their decision. More efficient decision makers might have pre-decided, which might account for the neural efficiency. The experiments lack any measurement of WHEN participants made their decision. For that reason, I would ask that the authors temper their claims about the significance of the alpha decoding and its possible causality.

   Thank you for this thoughtful comment. The two interpretations the reviewer raises—momentary brain state unconsciously biasing the upcoming choice versus subjects pre‑deciding during the inter‑trial period and holding that decision—are both, in principle, possible. Our design does not allow us to measure the exact moment at which each decision is made, so we cannot definitively rule out pre‑decision during the ITI. However, several considerations make a full “pre‑decide and hold” account less likely.

   First, choice trials constituted only one third of all trials. Any decision formed and held throughout the pre‑cue period would have to be discarded on the majority of trials when an instruction cue appeared. This would impose a relatively inefficient strategy that we did not encourage. Second, if subjects were consistently forming decisions well before the cue and maintaining them, we would expect above‑chance decoding from alpha patterns earlier in the pre‑cue interval. Instead, see Figure 4B, only alpha activity in the last ~500 ms prior to the cue reliably predicts the upcoming choice; earlier windows do not show robust predictive information. This temporal profile is more consistent with a decision process that crystallizes in close temporal proximity to the cue rather than being stably maintained over the entire 2–8 s interval.

   Regarding the broader temporal dynamics of the decision process, prior EEG work in this same paradigm (Rajan et al., 2018) showed that frontal theta activity specific to willed attention emerges 400–800 ms after the choice cue, and that the implementation of the posterior attention set (indexed by alpha lateralization) is delayed by approximately 400–500 ms in choice trials relative to instructed trials. Together, these findings place the main decision stage within the cue–target interval, rather than far in advance of the cue, and support the view that choices are formed spontaneously when the choice cue appears.

   At the same time, we fully agree that the present experiments do not provide a direct measure of when each individual subject commits to a particular choice, and that alpha decoding alone does not establish a causal influence of precue brain state on behavior. In the revised manuscript we therefore temper our wording about the implications of the precue alpha decoding, explicitly describing it as evidence that ongoing brain state carries predictive information about upcoming choices.

   (3) Did individual subjects exhibit a choice bias of location for the willed trials? If not, doesn't that raise concerns that subjects were trying to equalize their trials? If they do exhibit location biases, how does that impact the decoding? A simple decoder could learn to always just guess the biased direction for a subject and would perform > 50%. Consider the example in which an individual subject chooses 'Left' 55% of the time. A classifier that simply learns to choose 'Left' on every trial will be correct on 55% of trials. The training data would likely be sufficient to learn the direction of choice bias in each individual subject. So the classifiers could perform significantly above 50% without learning anything beyond the tendency of each subject. That is to say, 50% is not truly chance in this data set. It doesn't appear that Permutation testing has been performed to empirically determine chance for an individual's data. Permutation methods, scrambling the labels 1000 or 10000 times to establish a true baseline would be preferred over simply comparing to 50% and would address concerns about individual subject biases.

   Thank you for this thoughtful comment. To address the possibility that individual choice biases might inflate decoding above 50% accuracy, we conducted additional analyses.

   First, we examined whether subjects showed systematic left/right choice biases on willed trials. Figure A3A shows left/right choice proportions for each subject and dataset; dots indicate individual subjects’ percentages of left choices. Across the combined datasets, the mean proportion of left choices was 48.9%, which did not differ significantly from 50% (p = 0.581). At the individual-subject level, 10 of 31 subjects fell outside the 40–60% range, indicating that some degree of bias was present for a subset of participants, but that there was no consistent group-level bias toward one side.

   Second, to test whether such biases might drive the decoding results, we repeated the precue alpha decoding after equating the number of choice-left and choice-right trials within each subject. We then performed 10,000 within-subject label permutations to obtain an empirical chance distribution for each subject and for the group. In Figure A3B, balanced decoding of 52.81% was significantly above the empirical permutation baseline decoding accuracy of 50.01% at p = 0.028, indicating that precue decoding remains significantly above chance after accounting for trial imbalance and subject-specific choice biases. We now include these results in the Supplementary Materials and describe them in the revised manuscript, and we explicitly state that decoding accuracy is evaluated against an empirical permutation baseline rather than assuming a fixed 50% chance level.

   (4) The novel contributions of this work beyond the two prior Bengson et al papers from Dr. Mangun's lab appear quite modest. The discussion would be enhanced by specifically stating how the present work advances understanding beyond the prior Bengson studies.

   The Bengson studies have indeed laid an important foundation for the present work, and we agree that it is useful to explicitly articulate how our study extends their findings. Below we highlight three main advances.

   First, Bengson et al. (2015) showed that the choice cue in the willed attention paradigm activates a frontoparietal network, including regions such as dACC, DLPFC, and parietal cortex. In the present work, we replicate this activation pattern and go beyond it by using multivoxel pattern analysis to demonstrate that this network carries information about which spatial alternative is chosen (left vs right) on each willed attention trial. In other words, we attach the specific function of representing willed spatial decisions to this previously identified frontoparietal network, rather than only showing that it is engaged by the task.

   Second, Bengson et al. (2014) reported that choice-left and choice-right trials are accompanied by different patterns of precue alpha lateralization. Here, we build on that result by showing, with MVPA decoding, that distributed patterns of alpha power in the precue interval systematically predict the subsequent choice, and by mapping the time course of this effect. Our decoding time-course analysis indicates that predictive information is confined to the final ~600 ms before cue onset, rather than being present throughout the entire inter-trial interval. This timing supports the interpretation that the decision about where to attend is formed relatively close in time to the cue, rather than being pre-planned early in the ITI and simply held until needed.

   Third, we introduce a neural efficiency metric that links precue alpha decoding to postcue frontoparietal activation and decoding. This analysis shows that individuals with stronger precue alpha-based predictability of their choices tend to show more efficient (i.e., lower-amplitude yet still decodable) frontoparietal responses when implementing those choices. By relating ongoing brain state to subsequent decision-related network engagement, this provides a novel connection between spontaneous precue dynamics and the efficiency of the decision network that was not addressed in the earlier Bengson studies.

   Taken together, these contributions extend prior work by (i) demonstrating that the frontoparietal network encodes the content of willed spatial decisions, (ii) showing that precue alpha patterns in a specific late time window bias those decisions, and (iii) linking these precue patterns to individual differences in the efficiency of decision-related neural responses. We further highlight these points in the revised manuscript to clarify the added value of the present study over the earlier Bengson et al. reports.

   Reviewer #3 (Public review):

   Summary:

   This manuscript analyzes two independent datasets collected at different sites. Using the same willed-attention paradigm (instructional vs. choice cues) and combining fMRI and EEG analyses, the authors investigate how attentional direction is selected when no external instruction is provided. Their main claims are that the dorsal attention network is engaged by both cue types, whereas the choice cue additionally involves a frontoparietal decision network. Moreover, left-versus-right attentional decisions can be decoded in this decision network only on choice trials, and multichannel pre-stimulus alpha patterns predict the subsequent attentional choice. Finally, individuals with more predictive alpha patterns show greater neural efficiency in the decision network, i.e., higher decoding with lower BOLD activation.

   The question is worthwhile and the two-site design is a genuine strength. At the same time, several central inferences rely on decoding analyses for which the statistical testing and cross-validation structure are not described in enough detail to assess robustness. In addition, using a ratio-based neural-efficiency measure make the interpretation more fragile than it needs to be. With a focused revision that tightens inference around MVPA and clarifies a few methodological points, I think the paper could become substantially more convincing.

   Thank you for the succinct summary of our work and positive assessment. We address the concerns below.

   Strengths:

   The work extends previous willed attention studies by attempting to link pre-stimulus alpha pattern predictability to post-cue frontoparietal representations, and by testing reproducibility across two datasets. The conceptual advance beyond previous studies, e.g., Bengson et al. (2015), however, depends on how solid the decoding-based evidence is and whether alternative explanations are convincingly excluded. At present, the strength of support is limited mainly by incomplete reporting and/or controls for MVPA significance testing, as well as potential inflation of decoding estimates if folds are not independent of run structure. Concerns about statistical assessment of decoding accuracy are well documented in the literature (Combrisson & Jerbi, 2015).

   Thank you for raising concerns regarding the analysis in the paper. We address these concerns below.

   Weaknesses:

   (1) The manuscript describes the decoding pipeline for both fMRI and EEG MVPA. However, it does not clearly specify how "significantly above chance" is determined for the fMRI ROI decoding, nor how multiple comparisons across ROIs are handled, even though p-values are reported. The same issue applies to the time-resolved EEG analysis across many time points. For each decoding analysis, please specify the inferential test (e.g., permutation test within participant, group-level test on subject accuracies, binomial test, etc.) and report effect sizes with confidence intervals (e.g., Combrisson & Jerbi, 2015). Further, for EEG decoding over time, it would be preferable to control family-wise error, e.g., cluster-based permutation, rather than thresholding pointwise p-values. A standard approach here is the nonparametric cluster framework (e.g., Maris & Oostenveld, 2007).

   We thank the reviewer for this detailed and constructive comment. For the fMRI ROI decoding, each participant contributed one decoding accuracy value per ROI, obtained by averaging across cross-validation repetitions. In our original analysis, we tested whether decoding accuracy was significantly above the 50% chance level using group-level one-sample t-tests across participants, with FDR correction applied to control for multiple comparisons across ROIs. We now make this clear in the revised manuscript.

   Following the reviewer's suggestion, we additionally performed a random permutation test for each ROI, following the approach of Combrisson and Jerbi (2015), and computed effect sizes. See Table A1 (now included in the Supplementary Materials), decoding accuracy in all 9 ROIs was significantly above chance at the combined-dataset level, with multiple comparisons controlled via FDR. We include this permutation-based analysis in the Supplementary Materials and comment on the result in the revised manuscript.

   For the EEG time-resolved decoding, we followed the reviewer's suggestion and implemented a cluster-based permutation approach (Maris & Oostenveld, 2007) to control the family-wise error rate across time points. At each time point, decoding accuracy was compared against the 50% chance level using a one-sided dependent-samples t-test across participants. Consecutive time points exceeding an initial threshold of p < .05 were grouped into clusters, with cluster size defined as the number of consecutive significant time points. We then constructed a null distribution of the maximum cluster size using 10,000 sign-flip permutations and evaluated the observed cluster against this null distribution to control the familywise error rate. Using this approach, the cluster identified in Figure 4 was significant at pFWER = .008 at the combined-dataset level. We now report this cluster-based permutation result in the revised manuscript, along with the original analysis, in the revised manuscript.

   (2) The cross-validation approach used here is appreciated and appropriate in principle. However, random 10-fold splits across trials can inflate accuracy if training and test folds share run-specific noise, scanner drift, or autocorrelated structure. The manuscript should indicate whether folds were blocked by run or randomized across the entire session. In addition, please report the number of trials per condition after artifact rejection and after removing short ITIs for the long prestimulus epochs (−2500 ms to 0 ms) for each dataset in the section of EEG preprocessing. Similarly, please report how often participants chose left vs. right on choice trials, and whether balanced folds (or an equivalent balancing procedure) were used if needed.

   We thank the reviewer for this careful comment. To clarify, cross-validation folds were randomized across the entire session rather than blocked by run. We acknowledge that this approach could in principle allow training and test folds to share run-specific noise or scanner drift, which may inflate accuracy estimates relative to run-blocked cross-validation. In the revised manuscript, we note this as a limitation of our cross-validation scheme.

   We have also added the requested trial-count information to the EEG preprocessing section, reporting the number of trials per condition after artifact rejection and after excluding short-ITI trials for the long prestimulus epochs (−2500 to 0 ms), separately for each dataset.

   Regarding choice behavior, we now report the proportion of left versus right choices on choice trials for each dataset (see also our response to Reviewer 2's related comment on choice bias, where we found a group mean of 48.9% left choices, not significantly different from 50%). To address potential imbalance in individual subjects, we repeated the decoding analysis using trial-balanced folds (equating the number of left- and right-choice trials within each subject) and confirmed that the pattern of decoding results was consistent with our original analysis. This balanced-fold analysis is now reported in the revised manuscript.

   (3) Moreover, ROI definition is not sufficiently specified and independence should be clarified. The ROIs are defined based on peaks from the choice-instructed univariate contrast (Table 2) and then used for MVPA. First, are these ROIs defined as spheres around peaks or using anatomical masks? What radius or voxel count was used? This needs to be explicit. Second, I am concerned about circularity risk. Although choice-vs-instructed selection is not identical to left-vs-right decoding, ROI selection from the same dataset can still bias descriptive estimates and encourage overinterpretation if not carefully justified (Kriegeskorte et al., 2009). At minimum, the authors should explain why their selection criterion is independent of the decoded contrast under the null, and ideally provide a robustness check using either anatomical ROIs or independently defined ROIs, e.g., from prior literature or an atlas.

   We thank the reviewer for this comment. Regarding ROI definition, the ROIs were defined as 5-mm-radius spheres centered on peak voxels from the choice > instructed contrast; no anatomical masks were used. This is now stated explicitly in the revised Methods.

   Regarding the concern about circularity, our ROIs were indeed defined using the choice > instructed contrast from the same dataset, and left–right decoding was subsequently performed within these ROIs. As the reviewer notes, this raises a general concern about "double dipping" (Kriegeskorte et al., 2009) whenever the selection contrast is not independent of the statistic used in the selective analysis under the null hypothesis. In our design, the choice–instructed contrast was constructed by collapsing across left and right trials within each condition, and was therefore driven by overall differences in activation magnitude between willed and instructed attention, not by the left–right spatial distinction that is the target of the decoding analysis. Under the null hypothesis of no left–right information in the multivoxel patterns, the mean activation level used for ROI selection and the spatial pattern used for decoding are distinct quantities: a region could show a large choice–instructed activation difference while carrying no left–right information whatsoever, or vice versa. This argues that the two statistics are not mechanically coupled, reducing the risk that ROI selection based on the former systematically inflates decoding accuracy for the latter. We have clarified this logic explicitly in the revised Methods.

   Thanks for the comment. Regarding ROI definition, the ROIs were defined as spheres around contrast peaks, where the radius is 5 mm. No anatomical masks were used. This is now made clear in the revised manuscript. Regarding potential circularity in our ROI-based analyses, our ROIs were defined using the choice > instructed contrast from the same dataset, and we then performed left–right decoding within these ROIs. As the reviewer notes, this raises a general concern about “double dipping” (Kriegeskorte et al., 2009) if the selection contrast is not independent of the statistic used for the selective analysis under the null hypothesis. In our design, the choice–instructed contrast was constructed by collapsing across left and right trials within each condition and thus was sensitive to overall differences in activation between willed and instructed attention conditions, but not to the left–right distinction that is the target of the decoding analysis. Under the null hypothesis of no information about left vs. right in the multivoxel patterns, the choice–instructed contrast and the left–right decoding statistic are therefore orthogonal by construction, which reduces the risk that ROI selection based on the former will inflate the latter.

   (4) Using an index of neural efficiency is conceptually interesting. However, if the denominator, computed as the activation difference between choice and instructional conditions, is near zero or noisy, the ratio can become unstable. I would rather see a multivariate model that treats activation and decoding as separate dependent measures, or a latent-variable approach, than a single ratio.

   We appreciate the reviewer’s thoughtful comment. Our aim in using the neural efficiency metric was to provide an intuitively interpretable summary measure linking frontoparietal activation amplitude to decoding performance, but we agree that ratio measures can be problematic if not carefully evaluated. We also agree that more sophisticated multivariate or latent-variable approaches could provide a richer characterization of the joint relationships among activation, decoding, and behavior than a single ratio. Implementing and validating such models would, however, require additional analytic development that is beyond the scope of the present manuscript. We have therefore tempered our language when discussing neural efficiency, explicitly acknowledging the limitations of ratio-based measures and noting that future work could profitably apply multivariate or latent-variable modelling to more fully capture these relationships.

   Recommendations for the authors:

   Reviewer #1 (Recommendations for the authors):

   The authors should update Figure 2 to include a color bar representing t or z values and explicitly state the thresholding used for the maps. It is also essential that Figure 5B be updated to use identical scaling for both the UF and UCD subplots to ensure a transparent comparison of effect sizes across the two datasets. Regarding the statistical metrics, the authors should provide a sensitivity analysis or permutation test for the neural efficiency metric to ensure the ratio is stable and not driven by low variance BOLD denominators in the denominator of their calculation. A report of the univariate alpha asymmetry results is also recommended to validate that the multivariate patterns correspond to established raw alpha lateralization patterns in the posterior sensors.

   We thank the reviewer for these comments. Figure 2 has been updated to include a color bar indicating t-values, and the figure legend now explicitly states the statistical thresholding used to generate the map (including the threshold value and correction method applied). Figure 5B has also been revised so that the UF and UCD subplots now use identical axis scaling, allowing for a transparent, direct comparison of effect sizes across the two datasets.

   Regarding additional sensitivity analyses for the neural efficiency metric, our aim in using this metric was to provide an intuitively interpretable summary measure linking frontoparietal activation amplitude to decoding performance, and we agree that ratio measures can be problematic if not carefully evaluated. In line with Reviewer 3's related comment, we have tempered our interpretation of this metric in the revised manuscript and now explicitly describe it as an auxiliary, descriptive index rather than a primary inferential measure. We also note that more sophisticated multivariate or latent-variable approaches, as suggested by the reviewer, would provide a more rigorous characterization of the joint relationships among activation, decoding, and behavior, but that implementing such models is beyond the scope of the present study.

   Regarding univariate alpha asymmetry, Bengson et al. (2014) have already shown that the alpha asymmetry pattern prior to the choice cue in this paradigm is consistent with established attentional modulation of posterior alpha lateralization. We now explicitly note in the revised manuscript that our multivariate precue alpha decoding builds on this prior univariate finding, and we cite Bengson et al. (2014) as providing the corresponding raw alpha lateralization results.

   The methods section should be updated to explicitly state which TRs were used for the fMRI decoding and provide a justification for why a time-resolved approach was not utilized given the richness of the temporal data. Furthermore, the authors should provide a theoretical explanation for the observed behavioral parity between conditions. Specifically, it would be beneficial to discuss why the additional metabolic cost and recruitment of the willed network do not result in a behavioral tax such as slower reaction times or lower accuracy. This discussion should address whether the recruitment of the anterior insula and dACC might trigger an internal arousal shift or compensatory effort that allows the brain to maintain performance levels identical to the simpler instructed condition. Finally, the authors should clarify if site specific effects were modeled in their meta-analysis to account for the performance differences between UF and UC cohorts.

   We thank the reviewer for these helpful suggestions. With respect to the fMRI decoding, we did not use TR‑level data. Instead, we modeled each trial using a single‑trial GLM and extracted one beta estimate per trial, which served as the input to the multivoxel pattern analysis. In other words, the decoding was performed on trial‑wise parameter estimates rather than on individual TRs. We have clarified this explicitly in the Methods. Because our primary question about temporal dynamics was addressed using EEG, we did not implement a time‑resolved fMRI decoding approach. We now briefly justify this choice in the Methods, noting that the richer temporal information from EEG was used to characterize time courses, whereas the fMRI analysis focused on spatial patterns at the level of trial‑wise responses.

   Regarding the relation between behavior and insula/dACC activation, we agree that this is an important issue and have addressed it in detail in our response to Reviewer 1’s comment #2 in the Public View section above.

   Finally, we did not model site‑specific effects in our meta‑analysis. We now state this explicitly in the Methods.

   Reviewer #2 (Recommendations for the authors):

   Copying key text from the above section:

   (1) It would be helpful to discuss what other phenomena might fall within 'willed attention' and what some future studies might choose to focus on, along with some potential pitfalls

   (2) If possible, state your understanding of what processes where engaged by the decision component of the Willed Attention trials.

   (3) Discuss the timing of when subjects made decisions. Please temper claims about the significance of the alpha decoding and its possible causality.

   (4) Re-run MVPA analysis with Monte Carlo Methods to rule out potential biases at the level of individual subjects and to more properly establish chance levels. That is, scramble the labels and run it at least 1000 times per subject to establish what chance actually is.

   (5) The discussion would be enhanced by specifically stating how the present work advances understanding beyond the prior Bengson et al. studies.

   We appreciate the reviewer’s detailed suggestions and have addressed each of the points (1–5) in our responses above. In the revised manuscript, we now (1) expand the Discussion to consider other phenomena that may fall under willed attention and potential future directions and pitfalls, (2) clarify our interpretation of the processes engaged by the decision component of the willed attention trials, (3) discuss the timing of decisions and temper our claims regarding the significance and possible causality of the precue alpha decoding, (4) add permutation-based analyses of the MVPA to better characterize empirical chance levels and potential individual-subject biases, and (5) explicitly articulate how the present work extends the prior Bengson et al. studies.

   Minor Concerns:

   (1) Introduction. I found this set of statements to be confusing. I ‘Think’Table 1 that the statement implies that several one-second intervals were examined. If so, changing 'sequential 1 sec period' to 'sequential 1 sec periods' would help to reduce confusion. "The previous study (Bengson et al., 2014) reported that the amplitude of left parieto-occipital EEG alpha power in the 1 sec period immediately before the unpredictable appearance of the choice cue predicted the decision about where to attend. How far does this alpha-based predictability extend before the choice cue? Bengson and colleagues (2014) investigated sequential 1 sec period prior to the choice cue, and reported no significant predictive value of the parieto-occipital alpha power earlier in time.

   Thank you for pointing out this ambiguity. In the revised Introduction, we have clarified that Bengson et al. (2014) examined multiple sequential 1‑s periods prior to the choice cue, and we now use the plural “periods” to reflect this. We also rephrased the passage to more clearly distinguish their 1‑s windowed analysis from our finer‑grained time‑course decoding, which refines their result by localizing the predictive alpha pattern to approximately the final 600–700 ms before the cue.

   (2) Methods - Neural Efficiency: The definition of 'Neural efficiency' appears to be novel. This is quantified by the difference in decoding accuracy (choice vs. instructional) divided by the difference in BOLD activation (choice vs. instructional). Was this Neural Efficiency measure pre-registered or only derived post-hoc? While this is an intriguing construct, it is not clear why this metric was chosen over simply looking at the numerator. Isn't fMRI decoding accuracy the more relevant measure to compare to EEG decoding accuracy?

   We appreciate the reviewer’s careful comments on the neural efficiency metric. This measure was not preregistered; it was introduced post hoc as an intuitively interpretable, summary index relating frontoparietal activation amplitude to decoding performance. In defining it as the difference in decoding accuracy (choice vs. instructed) divided by the difference in BOLD activation (choice vs. instructed), our goal was to capture, in a single quantity, how much additional decodable information is obtained per unit increase in metabolic activity, in line with common informal uses of “neural efficiency” as doing more with less.

   We agree that fMRI decoding accuracy alone is a highly relevant and more standard measure to compare with EEG decoding accuracy, and we do not claim that the ratio is superior to examining the numerator by itself. In the revised manuscript, we therefore (i) clarify that the neural efficiency index was exploratory and not preregistered, (ii) present the decoding results themselves as the primary inferential measure, and (iii) describe the ratio explicitly as a descriptive, auxiliary metric that combines activation and decoding in a way that is conceptually appealing but statistically more fragile. We also note that more formal multivariate or latent‑variable approaches, as suggested by the reviewer, would provide a more principled way to model the joint relationships among activation, decoding, and behavior, but that developing such models is beyond the scope of the present study.

   (3) Methods - fMRI MVPA - which TRs from the trial were included in the MVPA analysis? All TRs following the onset of the cue? Does this include the actual deployment of spatial attention? Did any regions of the brain show effects for the instructed trials? One would think that even the univariate analysis would reveal right/left differences in dorsal attention network regions, including those in lateral frontal cortex.

   We thank the reviewer for these insightful questions. With respect to the MVPA analysis, we did not operate directly on TR‑level time series. Instead, we modeled each trial with a single‑trial GLM, using a canonical hemodynamic response function aligned to cue onset, and extracted one beta estimate per trial. These trial‑wise parameter estimates, which summarize the cue‑evoked BOLD response over the relevant post‑cue TRs, served as the input to the multivoxel pattern analysis. In other words, the decoding was performed on trial‑wise beta images rather than on individual TRs. Because these betas reflect the cue‑locked response, they capture activity associated with the deployment of spatial attention following the choice or instruction cue.

   Regarding effects in the instructed trials, the present work focused primarily on regions identified by the choice > instructed contrast, and on decoding within this willed attention network. The attention effects of instructed trials, including univariate contrasts between attending left and attending right in dorsal attention network regions and the visual cortex, are analyzed and reported in detail in Meyyappan et al. (2025). In the revised manuscript, we now clarify this focus in the Methods and refer explicitly to Meyyappan et al. (2025) for a treatment of lateralized attention effects in DAN and the visual system.

   (4) Results: It is curious that RTs differed by ~100 ms across data collection sites. Could the authors offer an explanation for this difference? This doesn't seem critical for any of the paper's claims, but it is a large effect left unexplained.

   The reviewer is correct to point out the quantitative differences in behavioral performance between the UF and UCD cohorts, with the UCD participants exhibiting slower reaction times and somewhat lower accuracy despite the use of the same willed‑attention paradigm and matched stimulus parameters. The exact source of these differences is not fully known, but one plausible contributor is procedural differences between sites: at UF, EEG and fMRI were recorded simultaneously in a single session, whereas at UCD, EEG and fMRI were acquired in separate sessions, which may have affected overall task engagement, familiarity with the task, and fatigue. We now comment on this site‑related behavioral difference in the revised manuscript.

   Reviewer #3 (Recommendations for the authors):

   Minors:

   (1) The sample characteristics are somewhat imbalanced by sex in the UCD dataset. I am not suggesting that a gender analysis is essential, but this should be transparently reported and briefly noted as a limitation.

   We appreciate the reviewer’s comment. In the revised manuscript, we now clearly point out the sex distribution of participants in the UCD dataset in the Methods section and explicitly note the resulting sex imbalance as a limitation.

   (2) Please clarify whether any site differences in acquisition, e.g., TR, slice order, EEG montage, simultaneous vs. separate EEG recording, were modeled as moderators.

   We did not model these factors as moderators. All datasets were acquired using highly similar acquisition protocols, and our primary analyses were conducted at the within-subject level, which reduces the impact of between-site differences. We now note this explicitly in the Methods.

   (3) In several places the writing uses strong mechanistic language, e.g., "These results demonstrate that the network of frontal and parietal brain regions preferentially activated during willed attention is involved in deciding (choosing) the direction of attention (left vs. right)." Given the correlational nature of the evidence, I would suggest toning this down to "consistent with" unless stronger causal constraints are added.

   The reviewer’s point is well taken. We have revised the manuscript to tone down mechanistic language, replacing strong causal statements with more conservative phrasing (e.g., “consistent with”) that reflects the correlational nature of our evidence.

   References:

   Rajan A, Siegel SN, Liu Y, Bengson L, Mangun GR, Ding M (2018). Theta Oscillations Index Frontal Decision-Making and Mediate Reciprocal Frontal–Parietal Interactions in Willed Attention. Cerebral Cortex 29(7): 2832-2843

   Nadra JG, Mangun GR (2023). Placing willed attention in context: a review of attention and free will. Frontiers in Cognition 2: https://doi.org/10.3389/fcogn.2023.1205618

   Bengson JJ, Kelley TA, Mangun G R (2015). The neural correlates of volitional attention: A combined fMRI and ERP study. Human Brain Mapping 36(7): 2443–2454.

   Bengson JJ, Kelley TA, Zhang X, Wang JL, Mangun GR (2014). Spontaneous Neural Fluctuations Predict Decisions to Attend. Journal of Cognitive Neuroscience 26(11): 2578– 2584.

   Liu Y, Hong X, Bengson JJ, Kelley TA, Ding M, Mangun GR (2017). Deciding where to attend: Large-scale network mechanisms underlying attention and intention revealed by graphtheoretic analysis. NeuroImage 157: 45–60.

   Kriegeskorte N, Simmons WK, Bellgowan PSF, Baker CI (2009). Circular analysis in systems neuroscience: The dangers of double dipping. Nature Neuroscience 12(5): 535–540.

   Combrisson E, Jerbi K (2015). Exceeding chance level by chance: The caveat of theoretical chance levels in brain signal classification and statistical assessment of decoding accuracy. Journal of Neuroscience Methods 250: 126–136.

   Maris E, Oostenveld R (2007). Nonparametric statistical testing of EEG- and MEG-data. Journal of Neuroscience Methods 164(1): 177–190.

   Read the original source
6. 

   Version published to 10.7554/elife.110199.2 on eLife

   Nov 10, 2026
7. 

   Version published to 10.7554/elife.110199 on eLife

   Nov 10, 2026
8. 

   Version published to 10.64898/2025.12.22.696009v2 on bioRxiv

   Aug 24, 2026
9. 

   eLife

   Mar 10, 2026

   eLife Assessment

   This useful study supplements previous publications of willed attention by addressing a frontoparietal network that supports internal goal generation. The evidence is solid in analyzing two datasets collected at different independent sites, using the same willed-attention paradigm and combining fMRI and EEG. This work will interest cognitive psychologists and neuroscience researchers.

   Read the original source
10. 

    eLife

    Mar 10, 2026

    Reviewer #1 (Public review):

    Summary:

    This study addresses a fundamental question in cognitive neuroscience regarding how the brain transitions from a reactive state of following external instructions to a proactive state of self-directed agency. The authors utilize an ambitious multimodal design by combining the spatial precision of fMRI with the temporal resolution of EEG across two independent datasets from the University of Florida and UC Davis. By applying multivariate pattern analysis, the work demonstrates that while both instructed and willed attention engage the Dorsal Attention Network, willed choices uniquely recruit a frontoparietal decision network including the dACC and anterior insula. Furthermore, the study shows that pre cue alpha oscillations can predict subsequent spontaneous choices. This provides a neural link …

    Reviewer #1 (Public review):

    Summary:

    This study addresses a fundamental question in cognitive neuroscience regarding how the brain transitions from a reactive state of following external instructions to a proactive state of self-directed agency. The authors utilize an ambitious multimodal design by combining the spatial precision of fMRI with the temporal resolution of EEG across two independent datasets from the University of Florida and UC Davis. By applying multivariate pattern analysis, the work demonstrates that while both instructed and willed attention engage the Dorsal Attention Network, willed choices uniquely recruit a frontoparietal decision network including the dACC and anterior insula. Furthermore, the study shows that pre cue alpha oscillations can predict subsequent spontaneous choices. This provides a neural link between pre-existing brain states and intentional action, representing a significant technical effort to characterize the neural scaffolding of internal goal generation.

    Strengths:

    The primary strengths of this work include the integration of fMRI and EEG which allows the authors to bridge the gap between slow metabolic signals and fast oscillatory brain states. The use of two independent cohorts is a commendable effort to ensure the reproducibility of the willed attention effect, which is often a concern in small sample neuroimaging studies. Additionally, the move beyond univariate activation toward information based mapping demonstrates that the identified networks actually contain specific information about the direction of attention.

    Weaknesses:

    However, several critical weaknesses must be addressed to support the fundamental claims made in the manuscript. There are significant behavioral differences in performance between the two sites, specifically regarding the UC Davis cohort exhibiting slower reaction times and lower accuracy compared to the UF group. These discrepancies suggest potential differences in subject populations or experimental environments that are not currently accounted for in the neural models. The fMRI analysis lacks temporal precision because the use of beta series regression collapses the complex BOLD response into a single estimate per trial. This loss of temporal information obscures the evolution of the decision process and makes it difficult to distinguish whether the identified patterns represent a truly spontaneous choice or a slow building pre planned strategy.

    Furthermore, the EEG decoding approach utilized the entire topography of electrodes rather than a biologically motivated posterior region of interest. Given that alpha mediated spatial attention is traditionally localized to parieto occipital sensors, using the full electrode set risks the inclusion of non neural artifacts such as micro saccades or muscle activity which can contaminate multivariate classifiers. The introduction of the neural efficiency metric also requires further validation as the current ratio is mathematically sensitive to small denominators in the BOLD contrast.

    Crucially, the manuscript does not address the physiological implications of recruiting additional frontoparietal networks when behavioral performance remains identical across conditions. The activation of the anterior insula and dACC is frequently associated with increased autonomic arousal and effort. If the willed condition requires more extensive neural scaffolding to reach the same behavioral output as the instructed condition, it raises the question of whether this internal decision process is accompanied by changes in arousal levels. The authors should consider whether the lack of a behavioral tax is due to a compensatory increase in arousal, which could be reflected in the EEG data or pupil diameter if recorded, and potentially also in the amplitude of BOLD activity, which is being masked by the neural efficiency metric. Without an account of how the brain balances this increased computational demand without impacting behavioral performance, the functional significance of the willed attention network remains partially obscured.

    Read the original source
11. 

    eLife

    Mar 10, 2026

    Reviewer #2 (Public review):

    Summary:

    This manuscript combines fMRI and EEG investigations performed at two research sites to examine 'willed' or volitional visuospatial attention, as contrasted with more standard cued (or 'instructed') visuospatial attention. The primary findings are: 1) willed attention (vs. instructed attention) drives additional cortical circuitry across a broad fronto-parietal network; 2) the direction of willed attention, but not instructed attention can be decoded from the pre-cue EEG data and from MVPA analysis of the trial-level fMRI data; and 3) the subjects with high EEG decoding also exhibited high neural efficiency (i.e., high decoding with low BOLD signal change) in the fMRI data. The methods and data analysis are generally sound, and these results appear solid. On the negative side, it is not made clear …

    Reviewer #2 (Public review):

    Summary:

    This manuscript combines fMRI and EEG investigations performed at two research sites to examine 'willed' or volitional visuospatial attention, as contrasted with more standard cued (or 'instructed') visuospatial attention. The primary findings are: 1) willed attention (vs. instructed attention) drives additional cortical circuitry across a broad fronto-parietal network; 2) the direction of willed attention, but not instructed attention can be decoded from the pre-cue EEG data and from MVPA analysis of the trial-level fMRI data; and 3) the subjects with high EEG decoding also exhibited high neural efficiency (i.e., high decoding with low BOLD signal change) in the fMRI data. The methods and data analysis are generally sound, and these results appear solid. On the negative side, it is not made clear how the present findings extend our understanding beyond prior published work from one of the senior authors. There are also three significant concerns regarding interpretation of the findings. One has to do with the causal interpretation of the pre-cue alpha EEG signal determining the direction of willed attention. The second concern is the degree to which the present research paradigm adequately examines 'willed attention.' The third is that the MVPA analysis is not sufficiently described, and Permutation testing needs to be done to validate these findings. Otherwise, this manuscript appears methodologically sound, but questions about interpretation may mute the potential impact.

    Strengths:

    The focus on willed attention attempts to move beyond some of the many limitations of standard laboratory investigations of attention.

    The shared paradigm across two modalities and two research sites demonstrates solid reproducibility, even though a few minor differences are observed across sites.

    Weaknesses:

    (1) There are concerns about this experimental paradigm carrying the banner of Willed Attention, because the application of 'Will' appears quite modest. Yes, extra brain activity is exhibited for this condition vs. its control, but do the cognitive processes isolated adequately stand in for 'Willed Attention?" Willed attention, as operationally defined here, appears to involve a simple decision process prior to the shifting of spatial attention. The cue is internally generated, but after that the rest of the attentional processes appear identical to standard externally cued visuospatial attention experiments. This self-generated cue process likely involves some sort of memory/history of the recently selected cues and then some random-ish selection between A and B. This appears very similar to asking the subject to guess whether a fair coin flip will be heads or tails on each trial. A mental 'coin flip' feels like a very weak version of 'will.' As a potential remedy, it would be helpful to discuss what other phenomena might fall within 'willed attention' and what some future studies might choose to focus on, along with some potential pitfalls (e.g., the reasons why the current study avoided more robust exemplars of will).

    (2) The manuscript is lacking a description of the decision processes used during the willed attention paradigm and is lacking evidence as to WHEN subjects made their willed decision. Both of these points are of major concern:

    (a) The authors state: "For willed attention, participants were explicitly told to avoid relying on any stereotypical strategies of generating decisions, such as always attending the same/opposite side they attended during the previous trial, as well as to avoid randomizing or equalizing their decisions to choose left or right across trials; prior studies found that decisions to explicitly randomize decisions might invoke additional working memory related processes (Spence & Frith, 1999)." Subjects were instructed NOT to apply a simple heuristic and NOT to randomize or try to equalize their decisions, but exactly HOW the subjects made their decisions is not at all clear. What options does that leave? How does this strategy avoid the working memory-related processes mentioned in the Spence & Frith, 1999 citation? The brain regions that comprise the network of interest (aka Frontoparietal Decision Network) are activated by a very broad range of visual cognitive tasks, including many working memory paradigms. The Anterior Insula and dACC nodes Salience Network often simply reflect task difficulty. Obviously, making a choice is more cognitively demanding than not making a choice. The present experiments do not distinguish functional roles between different regions of the Frontoparietal Decision Network. On the whole, the study does very little to isolate the cognitive processes or neural bases of willed attention beyond calling out the set of 'Usual Suspects' for visual cognition.

    (b) The finding that pre-cue EEG signals predicted the postcue decision is intriguing. It could mean that the seemingly irrelevant and transient state of the brain causally and unconsciously biased the subject to one direction or the other. Alternatively, it could mean that the subjects utilized the pre-cue period to make their decision and hold it in case it was needed (i.e., that it was a choice trial). While 2-8 seconds ITI variability makes sense for fMRI decoding, it is a long time for a subject to idly wait, so they might fill that time preparing for the next trial. There appears to have been a substantial amount of individual difference in the pre-cue alpha decoding, which could reflect individual differences in cognitive strategy, specifically in the use of the pre-cue period to make their decision. More efficient decision makers might have pre-decided, which might account for the neural efficiency. The experiments lack any measurement of WHEN participants made their decision. For that reason, I would ask that the authors temper their claims about the significance of the alpha decoding and its possible causality.

    (3) Did individual subjects exhibit a choice bias of location for the willed trials? If not, doesn't that raise concerns that subjects were trying to equalize their trials? If they do exhibit location biases, how does that impact the decoding? A simple decoder could learn to always just guess the biased direction for a subject and would perform > 50%. Consider the example in which an individual subject chooses 'Left' 55% of the time. A classifier that simply learns to choose 'Left' on every trial will be correct on 55% of trials. The training data would likely be sufficient to learn the direction of choice bias in each individual subject. So the classifiers could perform significantly above 50% without learning anything beyond the tendency of each subject. That is to say, 50% is not truly chance in this data set. It doesn't appear that Permutation testing has been performed to empirically determine chance for an individual's data. Permutation methods, scrambling the labels 1000 or 10000 times to establish a true baseline would be preferred over simply comparing to 50% and would address concerns about individual subject biases.

    (4) The novel contributions of this work beyond the two prior Bengson et al papers from Dr. Mangun's lab appear quite modest. The discussion would be enhanced by specifically stating how the present work advances understanding beyond the prior Bengson studies.

    Read the original source
12. 

    eLife

    Mar 10, 2026

    Reviewer #3 (Public review):

    Summary:

    This manuscript analyzes two independent datasets collected at different sites. Using the same willed-attention paradigm (instructional vs. choice cues) and combining fMRI and EEG analyses, the authors investigate how attentional direction is selected when no external instruction is provided. Their main claims are that the dorsal attention network is engaged by both cue types, whereas the choice cue additionally involves a frontoparietal decision network. Moreover, left-versus-right attentional decisions can be decoded in this decision network only on choice trials, and multichannel pre-stimulus alpha patterns predict the subsequent attentional choice. Finally, individuals with more predictive alpha patterns show greater neural efficiency in the decision network, i.e., higher decoding with lower …

    Reviewer #3 (Public review):

    Summary:

    This manuscript analyzes two independent datasets collected at different sites. Using the same willed-attention paradigm (instructional vs. choice cues) and combining fMRI and EEG analyses, the authors investigate how attentional direction is selected when no external instruction is provided. Their main claims are that the dorsal attention network is engaged by both cue types, whereas the choice cue additionally involves a frontoparietal decision network. Moreover, left-versus-right attentional decisions can be decoded in this decision network only on choice trials, and multichannel pre-stimulus alpha patterns predict the subsequent attentional choice. Finally, individuals with more predictive alpha patterns show greater neural efficiency in the decision network, i.e., higher decoding with lower BOLD activation.

    The question is worthwhile and the two-site design is a genuine strength. At the same time, several central inferences rely on decoding analyses for which the statistical testing and cross-validation structure are not described in enough detail to assess robustness. In addition, using a ratio-based neural-efficiency measure make the interpretation more fragile than it needs to be. With a focused revision that tightens inference around MVPA and clarifies a few methodological points, I think the paper could become substantially more convincing.

    Strengths:

    The work extends previous willed attention studies by attempting to link pre-stimulus alpha pattern predictability to post-cue frontoparietal representations, and by testing reproducibility across two datasets. The conceptual advance beyond previous studies, e.g., Bengson et al. (2015), however, depends on how solid the decoding-based evidence is and whether alternative explanations are convincingly excluded. At present, the strength of support is limited mainly by incomplete reporting and/or controls for MVPA significance testing, as well as potential inflation of decoding estimates if folds are not independent of run structure. Concerns about statistical assessment of decoding accuracy are well documented in the literature (Combrisson & Jerbi, 2015).

    Weaknesses:

    (1) The manuscript describes the decoding pipeline for both fMRI and EEG MVPA. However, it does not clearly specify how "significantly above chance" is determined for the fMRI ROI decoding, nor how multiple comparisons across ROIs are handled, even though p-values are reported. The same issue applies to the time-resolved EEG analysis across many time points. For each decoding analysis, please specify the inferential test (e.g., permutation test within participant, group-level test on subject accuracies, binomial test, etc.) and report effect sizes with confidence intervals (e.g., Combrisson & Jerbi, 2015). Further, for EEG decoding over time, it would be preferable to control family-wise error, e.g., cluster-based permutation, rather than thresholding pointwise p-values. A standard approach here is the nonparametric cluster framework (e.g., Maris & Oostenveld, 2007).

    (2) The cross-validation approach used here is appreciated and appropriate in principle. However, random 10-fold splits across trials can inflate accuracy if training and test folds share run-specific noise, scanner drift, or autocorrelated structure. The manuscript should indicate whether folds were blocked by run or randomized across the entire session. In addition, please report the number of trials per condition after artifact rejection and after removing short ITIs for the long prestimulus epochs (−2500 ms to 0 ms) for each dataset in the section of EEG preprocessing. Similarly, please report how often participants chose left vs. right on choice trials, and whether balanced folds (or an equivalent balancing procedure) were used if needed.

    (3) Moreover, ROI definition is not sufficiently specified and independence should be clarified. The ROIs are defined based on peaks from the choice-instructed univariate contrast (Table 2) and then used for MVPA. First, are these ROIs defined as spheres around peaks or using anatomical masks? What radius or voxel count was used? This needs to be explicit. Second, I am concerned about circularity risk. Although choice-vs-instructed selection is not identical to left-vs-right decoding, ROI selection from the same dataset can still bias descriptive estimates and encourages overinterpretation if not carefully justified (Kriegeskorte et al., 2009). At minimum, the authors should explain why their selection criterion is independent of the decoded contrast under the null, and ideally provide a robustness check using either anatomical ROIs or independently defined ROIs, e.g., from prior literature or an atlas.

    (4) Using an index of neural efficiency is conceptually interesting. However, if the denominator, computed as the activation difference between choice and instructional conditions, is near zero or noisy, the ratio can become unstable. I would rather see a multivariate model that treats activation and decoding as separate dependent measures, or a latent-variable approach, than a single ratio.

    Read the original source
13. 

    Version published to 10.7554/elife.110199.1 on eLife

    Mar 10, 2026
14. Mentioned by @biorxiv-neursci.bsky.social on Bluesky

    Dec 24, 2025

    Neural Mechanisms of Willed Attention Control https://www.biorxiv.org/content/10.64898/2025.12.22.696009v1

15. Mentioned by @biorxivpreprint.bsky.social on Bluesky

    Dec 24, 2025

    Neural Mechanisms of Willed Attention Control https://www.biorxiv.org/content/10.64898/2025.12.22.696009v1

16. 

    Version published to 10.64898/2025.12.22.696009v1 on bioRxiv

    Dec 24, 2025