Introduction

Humans have an impressive ability to plan. We are able to model the world, simulate potential outcomes, and select among possible courses of action. Take, for example, your first trip to London. You want to visit Buckingham Palace despite being jet-lagged. Looking at a map of the underground, you’re overwhelmed with information but need to make a plan. How do you solve this problem? Even simple decisions like this involve many potential actions and outcomes, making it impossible to systematically evaluate every possible option, especially given limited cognitive resources^1–6. Explaining how people plan efficiently and flexibly under these constraints is a long-standing challenge in human and machine intelligence^5–8.

Theories of human problem-solving conceptualize planning as a search through a ‘decision tree’ of all potential actions and their outcomes^1–3,9,10. In our example, an individual may first list all possible tube stations within walking distance and then evaluate which action sequence will get them closer to their destination. Previous work proposes different algorithmic strategies for how an agent efficiently searches over a complex decision tree. These strategies include ignoring low-value actions (i.e., pruning)^3,11–13, limiting how far in the future one might search (i.e., depth)^1,14–16, or relying on previously learnt strategies (i.e., habits)^4,6,14,17.

This previous work, however, largely assumes that a decision-maker has a fixed representation of the problem. When planning involves constructing and evaluating multiple multi-step trajectories within a decision tree, the computational burden increases with the complexity of the representation of the problem space. Consider planning in a two-dimensional spatial grid, for example. A fine-grained grid presents many choice points about which way to turn. A coarser-grained grid presents fewer choice points. Since the number of branches is a multiplicative function of the number of choice points, a simplified representation of the task space, if chosen appropriately, can have a profound effect on reducing the computational demands of planning.

One elegant approach to forming such a simplified representation is to adaptively select the granularity of information required to successfully complete the task¹⁸, known as value-guided construal (VGC). Under VGC, a cognitively limited decision-maker selects a manageable subset of information over which to plan—i.e., a task representation— balancing utility and complexity¹⁸. In our example, the VGC algorithm predicts that an individual would plan over a few relevant tube lines rather than planning over all possible routes.

In previous work, Ho and colleagues discovered that people’s awareness of, and memory for, obstacles in a maze varies in line with the predictions of a VGC model. The VGC model implies two nested optimisations - an outer loop of construal, and an inner loop that runs a plan conditional on a particular task representation. The VGC model is a normative model and remains agnostic as to the cognitive mechanisms controlling the construal. In particular, the perceptual and attentional mechanisms governing how information is selected to become part of a task representation remain unknown. Initiating such a nested computation plausibly rests on inductive biases - general principles that a perceptual system can apply to select task-relevant information, before refining it as part of the planning process ¹⁹. Selective attention is proposed as one general mechanism by which the brain selects relevant information, either voluntarily (endogenous) or reflexively (exogenous)^20–25.

Previous studies have demonstrated that attention guides the selection of particular features of the environment to support reinforcement learning²⁶. However, it remains unknown whether and how attention shapes value-guided construal “on the fly” during planning. For instance, one possibility is that forming a simplified task representation is a “late” passive side-effect of the planning process - a tendency to focus on what we are thinking about. Alternatively, VGC may reflect an “early” selection of perceptual information, perhaps based on a rapid feedforward sweep of perceptual input²⁷. These alternatives echo classic debates between early and late selection models of attention^28,29, but now situated within the broader landscape of computational accounts of planning. More generally, despite the wealth of literature on attention, and pioneering efforts to incorporate attentional constraints into models of decision-making^18,26, we lack a basic understanding of how attention influences planning.

To make progress on this question, we examined the role of visuospatial attention on how people construct simplified task representations across three experiments in human participants. We build on previous work using maze navigation to provide a rich readout of people’s current task representations. We predicted that if visuospatial attention is guiding the formation of task representations, the construal process will be constrained by inductive biases characteristic of attentional selection. For instance, previous work has illustrated how attentional selection is biased by the spatial context in which information is presented: presenting distractors alongside task-relevant stimuli makes attentional selection more challenging^30–32. Attention, in this case, spills over to the neighbouring stimuli. These findings align with the metaphorical attentional spotlight, which stipulates that the focus of visual attention can move around the visual field but is limited in spatial extent^33,34. According to this model, individuals can, for example, orient their attention preferentially to a single hemifield—i.e., lateralizing—which is enabled by a hemispheric lateralization of alpha power over posterior cortex^35–38.

We harness these classical signatures of attentional selection to characterise how attention shapes planning. First, we demonstrate “attentional overspill”: participants preferentially incorporate task-irrelevant information into their task representation when it is presented in spatial proximity to task-relevant information. Second, we observe that attentional overspill is reduced when task-relevant information is lateralised to a single hemifield, allowing participants to more effectively form optimal task representations. Finally, we extend the VGC model to incorporate visuospatial attention as a key psychological mechanism for constructing simplified task representations. Together, our findings furnish a computational account of how attention and perception guide simplified representations in the service of planning.

Results

To examine the role of visuospatial attention in planning, we relied on a previously developed maze navigation paradigm in which participants solved 2-D mazes¹⁸, avoiding obstacles obstructing their path (Figure 1a, left panel). On every trial, participants reported their awareness of specific obstacles (see Methods for details). The level of awareness attached to different obstacles provides a read-out of an individual’s task representation while solving a particular maze. We first reanalyzed the data presented by Ho and colleagues (2022)¹⁸ to examine the role of spatial attention in building task representations (datasets Ho 1 and 2). In a new experiment (dataset dSC 1), we designed novel mazes to test the effects of lateralization of attention in enabling efficient planning (see Methods & Table S1).

Fig. 1

A spotlight of attention influences task representations

We hypothesized that spatial attention would control which items are included in a task representation^30–32. Specifically, we hypothesised that participants would deviate from the predictions of the VGC model and become distracted by task-irrelevant obstacles when they are presented in spatial proximity to task-relevant obstacles. To evaluate these predictions, we first computed the distance between a probe obstacle and every other obstacle in the maze. Second, we ranked the obstacles from the closest to the furthest from the probed item. Using the ranked obstacles, we trained a linear regression model to predict participants’ awareness of the probed obstacle (in green) from their awareness of the remaining obstacles (in grey; Figure 1b).

Critically, we observed a significant effect of spatial context on task representations - an effect which is not predicted by the normative VGC model. Participants’ awareness of a particular obstacle was positively predicted by the awareness of its close neighbours (Pi = 0.26, SE = 0.01,95% CI [0.25, 0.28]; β₂ = 0.29, SE = 0.01,95% CI [0.27, 0.30]), whereas awareness of its furthest neighbours negatively predicted participant reports (β₅ = −0.13, SE = 0.01, 95% CI [-0.15, −0.12]; β₆ = −0.13, SE = 0.01, 95% CI [-0.15, −0.12]; see Table S2). In other words, the spatial context of an obstacle predicted whether it would be included in a simplified task representation – akin to a diffuse attentional spotlight which filters which aspects of the maze are available for planning. This effect remained significant for both task-relevant and task-irrelevant obstacles, and after controlling for the predictions of the VGC model (Figure S7 & Table S3, respectively). We observed the same effect in a separate experiment where participants planned their route upfront before navigating the mazes (i.e., dataset Ho 2, see Table S4 & S5). Finally, we replicated this pattern of results in our in-person experiment: closest neighbours positively predicted the awareness of an obstacle (β₁ = 0.19, SE = 0.007, 95% CI [0.18, 0.21]), whereas furthest neighbours negatively predicted participants’ reports (β₃ = −0.10, SE = 0.01, 95% CI [-0.11, −0.08]; β₄ = −0.26, SE = 0.007, 95% CI [-0.27, −0.25]; β₅ = −0.29, SE = 0.007, 95% CI [-0.30, −0.27]; see Table S6 & S7 and Figure S11).

Next, we explored whether the influence of neighbouring obstacles on task representations varied across individuals. To do so, we fit the regression model described above to quantify each participant’s attentional spillover, and quantified the linear slope of the resulting beta coefficients. Negative slopes indicate a significant effect of attentional spillover on task representation. The influence of attention varied considerably across participants: while on average, participants’ task representations were influenced by attention (mean effect = −0.08; s.d. = 0.04), a subset of participants showed minimal influence of attention on their task representation (i.e., flat slopes; Figure 1c).

We hypothesized that participants with the largest attention effects (i.e., most negative slopes) would also show sparser task representation (i.e., a “spotlight of attention” which is focused only on a subset of obstacles). To test this, we computed the sparsity of participants’ task representations by estimating the variance of their awareness reports, with higher variance indexing those participants who report being very aware of some obstacles and unaware of others. In line with our hypothesis, we observed that participants who were most influenced by neighbouring obstacles also showed sparser task representations (dataset Ho 1: ρ = −0.35, p< 0.001 ; dataset Ho 2: ρ = −0.49, p< 0.001 ; dataset dSC 1: ρ = −0.51, p< 0.01; see Figure S11). To address concerns of overfitting, we tested whether the spatial attention effects observed in a lateralized set of mazes generalized to task representations of non-lateralized mazes and vice versa (dataset dSC 1). We observed that inter-individual differences in spatial attention effects in one condition predicted the sparsity of task representations in the other (ρ = −0.48, p< 0.01; ρ = −0.42, p< 0.05).

Attentional limits constrain the optimality of task representations

Prior psychological research indicates that attention can be efficiently allocated to a “hemifield” of visual space - with information being preferentially processed when presented in the attended hemifield^39–41. Building on this work, we hypothesized that participants would select task-relevant information with greater ease - constructing task representations more closely aligned with the VGC model - when task-relevant information is spatially confined to a visual hemifield (i.e., presented unilaterally).

To test this hypothesis, we derived a measure of task-relevant lateralization inspired by the attention literature^35,42,43 (Figure 2a). Specifically, we separated maze stimuli across the vertical meridian and computed the ratio of task-relevant information presented on the left versus right side. For example, the maze shown in Figure 2a has twice the amount of task-relevant information presented in the left hemifield than in the right (lat. Index= 1/3). A lateralization index of 0.0 indicates that both hemifields contain equal amounts of taskrelevant information (i.e., non-lateralized). We used this task-relevance lateralization index as a moderator in a hierarchical linear regression model to test whether participants’ awareness reports were better predicted by the original VGC model in mazes showing the greatest lateralization of task-relevant information.

Fig. 2

In line with our hypothesis, we observed a significant moderation effect whereby the greater the lateralization of task-relevant information across the vertical meridian, the better the original VGC model was at predicting participants’ awareness reports (β_interaction = 0.01, SE = 2.65*10⁻³, 95% CI [0.01, 0.02], ρ_FDR< 0.001; Figure 2c left panel & Table S8). We replicated these findings with the data collected in dataset Ho 2¹⁸ (ρ_FDR< 0.01; see Table S9). These results indicate that participants’ task representations are more closely aligned with the ideal observer (i.e., the original-VGC model) when task-relevant information is presented unilaterally.

In our new dataset (dSC 1), we designed novel maze stimuli to validate these lateralised effects of attention while addressing some limitations of previous experiments (see Methods). We again observed that lateralization of task-relevant information impacted participants’ awareness reports. Participants were less aware of task-irrelevant stimuli on trials where the lateralization of task-relevant information was larger (Figure 2b) and we replicated the moderation effect of information lateralization on the extent to which the original VGC model captured participants’ awareness reports (β_interaction = 0.01, SE = 2.65*10⁻³, 95% CI [0.01, 0.02], p< 0.001; Figure 2c & Table S10).

In contrast with our observations of consistent and strong attentional effects relative to the vertical meridian, effects relative to the horizontal meridian (superior vs. inferior) were inconsistent across experiments. Specifically, we observed a significant moderation effect in dataset Ho 2 (α_interaction = 0.01, SE = 2.85*10⁻³, 95% CI [0.00, 0.01], p^FDR< 0.05; see Table S9), but not in dataset Ho 1, and the moderation effect was negative rather than positive in dataset dSC 1 (β_interaction = −0.01, SE = 2.22*10⁻³, 95% CI [-0.01, 0.00], p< 0.05). Both of these effects became insignificant after accounting for nuisance covariates (see Table S14 & S15).

Attentional spotlight model of task representations

Taken together, our results corroborate a critical role for visuospatial attention in constructing task representations. Notably, these filtering effects of attention on value-guided construal are not currently part of the original VGC framework proposed by Ho and colleagues. In what follows we explicitly incorporate the influence of a spotlight of attention into the original VGC model to formulate the spotlight-VGC model^39,40.

To achieve this, we computed the predictions of the existing VGC model for each obstacle’s task relevance in a given maze, and averaged these predictions within an attentional spotlight of 3 squares (Figure 3a & S8, see Methods for details). We depict the effects of this spatial spotlight in Figure 3a: task-irrelevant stimuli (plotted in grey; see middle left obstacle) neighbouring task-relevant obstacles (plotted in orange) become more task-relevant, whereas task-relevant information becomes less relevant when surrounded by task-irrelevant information (see bottom right orange obstacle). We hypothesized that this spotlight-VGC model would predict participants’ reports better than the original VGC model, which does not account for spatial attention.

Fig. 3

In line with this hypothesis, we observed that the spotlight-VGC model predicted participants’ awareness reports better than the original VGC model in all three datasets (dataset Ho 1: ΔBIC= 84.63; Ho 2: ΔBIC= 203.43; dSC 1: ΔBIC= 70.72; see Figure 3b right panel). For dataset dSC 1, we observed a significant improvement in model fit for non-lateralized maze stimuli (ΔBIC= 161.93) but failed to find any improvement when maze stimuli were lateralized (ΔBIC= −42.02; see Figure 3c). These findings dovetail with the previously discussed moderation effects, and suggest that the spotlight-VGC model is particularly useful in improving our ability to explain human behaviour in situations when attentional filtering is more complex.

Sensitivity analyses

We conducted a series of control analyses to verify the robustness of our experimental results. First, we verified that the spatial proximity effect (Figure 1b) was not driven solely by the spatial smoothness of participants’ awareness reports by conducting null permutation tests (grey line, Figure 1b). For each maze stimulus, we permuted the rank of the neighbouring obstacles. We then fit the same linear model to assess the effect of spatial context on task representations. This procedure was repeated 1000 times to generate a null distribution of beta coefficients. The resulting null distribution showed no discernible effect of spatial context. Second, we used null permutation tests to verify that the improved fit of the spotlight-VGC model was not driven by greater spatial smoothness of the model predictions (see Supplemental Figure 12). Third, we assessed whether nuisance covariates could explain the moderation effects we observed. Specifically, we added the distance from the goal, starting location, center walls, and fixation as nuisance covariates in our hierarchical regression models. Maze lateralization remained a significant moderator of the relationship between the original VGC model and participants’ awareness reports after controlling for these covariates (see Table S11–13). This was not the case, however, for lateralization effects along the horizontal meridian (see Table S14–15).

Finally, we sought to verify that the lateralization effects we observed were not driven by a change in eye movement patterns. For dataset dSC 1, we continuously tracked the position of participants’ gaze. We explicitly instructed participants to maintain central fixation while planning (see Methods for details) and removed the obstacles from the screen after 6 seconds. This allowed us to verify that greater awareness of obstacles was not driven by longer fixation times. We confirmed that participants maintained central fixation on both lateralized and non-lateralized maze stimuli in most trials (see Figure S13). Excluding trials where participants exhibited excessive eye movements during planning, we continued to observe qualitatively similar lateralization effects (see Figure S14 and Table S16).

Discussion

Searching for a solution in a complex multi-step task is challenging. Recent computational work suggests that humans overcome this challenge by constructing simplified perceptual representations of their environment. In the present study, we reveal a role for visuospatial attention in constructing these simplified perceptual representations.

Participants’ task-representations are informed by visuospatial attention

We provide several lines of evidence for the critical role of visuospatial attention in constructing task representations. First, we observed a significant effect of the spatial context in which information is presented. Participants were less likely to incorporate taskrelevant information into their construal when it was surrounded by task-irrelevant information. These effects mirror perceptual crowding effects^30,32 which reveal that attention spills over to distractors presented alongside task-relevant stimuli when presented in close spatial proximity. Second, we observed that participants incorporated task-relevant information into their task representations more frequently when relevant obstacles were grouped together within the same hemifield (Figure 2). Participants’ task representations in such settings were more closely aligned with an ideal observer model - suggesting that the natural contours of visuospatial attention interact with the capacity of observers to form efficient task representations.

We also observed significant inter-individual differences in attentional effects across participants (Figure 1c). While some participants were strongly influenced by the spatial context of neighbouring stimuli, others showed more limited evidence for an attentional effect (Figure 1b). Inter-individual differences in attention predicted the sparsity of participants’ simplified representations: participants with larger attention effects exhibited sparser representations. Future research could explore how these individual differences constrain performance on other tasks that require planning and search in highdimensional spaces.

Incorporating attention into a model of value-guided construal

The VGC model articulates how an ideal decision-maker should represent an environment while balancing complexity and utility¹⁸. We developed an extension of this model that accounts for the effects of spatial attention on planning. Our model, inspired by the analogy of a spotlight of attention^23,39, provides a better fit to participants’ awareness reports than the original VGC model (Figure 3). This improved model fit was most evident for mazes where task-relevant information was presented to both hemifields (Figure 3c), suggesting the augmented model is helpful in explaining behaviour in contexts where attentional selection is more complex. These deviations from the original VGC model, therefore provide a useful benchmark to compare human performance and offer insights into natural constraints on human cognition. For instance, we demonstrate that spatial context biases whether information is to be included or excluded from a representation of the environment. These effects may reflect inductive biases in humans who have learned and evolved to select information from real-world environments where obstacles tend to be grouped together in visual scenes^44,45. However, it is plausible that these inductive biases on value-guided construal may themselves be learnt, and vary according to other environmental demands and contexts which impose systematic regularities on useful task representations (e.g., attending preferentially to intersections when planning on the Tube). Future research can explore the flexibility of participants’ task representations across environmental contexts, and ask how these inductive biases are acquired.

Planning and Consciousness

Our experiments investigate the connection between planning and participants’ reports of their awareness of features of the task environment. The results may therefore be relevant to understanding the functions of conscious experience. While an intimate connection between attention and consciousness is widely recognized^46–49, there is less work explicitly considering the connection between planning and consciousness^50,51. However, there are several reasons why the kind of planning at work in our experiments is likely to require the task to be represented consciously.

As we mentioned at the outset, simplified representations reduce the computational burden of planning in a branching multi-step task space. The same consideration suggests that planning should be based on conscious, rather than unconscious, representations⁵². Initial stages of perceptual processing can carry information about a range of different and incompatible possibilities at once, for example, a probability distribution across a range of possible orientations of a line. The probabilistic representation attaches some probability (or probability density) to many different possibilities. There is, of course, a certain burden in integrating and weighing probabilistic information of this kind, for which the brain is thought to deploy various solutions (e.g. approximate Bayesian inference). These initial stages of perceptual inference are typically thought to be unconscious. However, forward planning from multiple possibilities in a branching task space rapidly becomes intractable as the combinatorial possibilities explode. Consciousness, by contrast, provides a much sharper representation of the current state, from which planning can proceed forward^53,54.

Given the computational cost of running through and comparing many potential multi-step action sequences, it makes sense to base that process on a reliable estimate of the current world state. While it is doubtless useful to produce some kinds of unlearned and habitual action very rapidly at the first hint of information, for example of the presence of a predator, with multi-step forward planning it makes sense to integrate information from more sensory modalities and across a longer timescale before then committing to using a representation as the basis for planning. This again suggests that conscious representations, which are known to integrate information across modalities and time^55–59, are perfectly suited to the functional needs of this kind of planning task. Furthermore, planning depends on both facts and values. Potential actions are assessed based on the expected value of outcomes. The role of value was captured, in our studies, by an extension of the VGC (value-guided construal) model. Consciousness is thought to facilitate the integration of different sources of value^60–62.

The task and associated computational model thus offer a flexible tool for characterising the computations by which conscious representations influence decisions and actions. Future work could tell us more about the way bottom-up attention-driven inputs and taskbased value jointly influence what information reaches conscious experience. This provides a novel ecologically-valid probe of the connections between attention, consciousness, and decision-making that does not require the explicit labelling of taskrelevant stimuli. Neural (e.g. M/EEG) data collected while participants plan could help understand the timescale and computational steps that lead to the formation of a conscious task representation. Modifications of the paradigm would also be suited to exposing the way non-consciously-presented cues do and do not influence the way participants plan.

Methodological Considerations and Future Directions

The task we used requires participants to report their experience of each maze. Yet, retrospective reports are one step removed from perception and do not shed light on the time course of how perceptual representations are formed. In line with this limitation, the extended spotlight-VGC model accounts for visuospatial attention effects after a simplified representation has been computed (i.e., at the end of the construal process). This allowed us to evaluate the fit of the model to empirical data, but does not provide insight into how attention affects the construal process itself. Future research examining the neural dynamics of task representations will be critical to disentangling the time course of how attention and planning considerations interact in the formation of task representations. It will also be necessary to elaborate on how bottom-up and top-down aspects of attentional selection are combined to guide complex behaviours. Finally, open questions remain about the nature and form of visuospatial attention and how this intersects with planning. For instance, can multiple spatial locations be preferentially selected at once—i.e., are there multiple spotlights^63–66? There is also discourse on how spatial attention may move from one location to another: are the intervening visual regions between attended locations similarly selected^63,67–69? Our findings tentatively suggest that individuals are able to attend to disparate spatial regions to form sparse task representations, yet there is substantial variability in how individuals orient their attention during the task. The present paradigm and computational modelling, in conjunction with carefully designed stimuli, may help resolve these outstanding questions.

While we observed clear lateralization effects along the vertical meridian (i.e., left vs right hemifield), effects along the horizontal meridian were less clear (see Table S14–15). A combination of factors may explain this finding. First and foremost, the maze stimuli were not designed to test horizontal-meridian lateralization effects, possibly leading to a lack of power to detect these effects. Second, prior research suggests that distractors produce a larger crowding effect when presented across the horizontal compared to the vertical meridian³⁰. The retinotopic organization of the cortex is believed to drive this effect: spatially adjacent stimuli can be retinotopically distant if presented on the opposite side of the vertical meridian, which is thought to facilitate distractor inhibition. A more detailed experiment explicitly designed to test these effects may clarify the interpretation of these null results.

Conclusions

Complex daily decisions require a decision-maker to arbitrate over countless potential multi-step actions and their outcomes, making searching for a solution difficult. We shed light on how this is achieved by clarifying the role of visuospatial attention in forming simplified perceptual representations to aid in planning. We build on previous work on the effect of task relevance and develop a computational model which explicitly incorporates the role of attention in value-guided construal. Our model bridges the literature on perception, attention and computational models of planning to provide a more complete computational account of human cognition. We believe the results of this paper can inform future research on a comprehensive theory of human cognition and inspire novel biologically-informed intelligent algorithms.

Methods

Experimental Task

To test our hypotheses we relied on a previously established mazenavigation task where participants are asked to move a circle avatar from a starting location to a goal using the arrow keys¹⁸. Each maze consisted of an 11x11 grid with blue obstacles (7 obstacles in datasets Ho 1 & 2, and 6 obstacles in dataset dSC 1), and black central walls arranged in the shape of a fixation cross. Each trial began with a fixation cross (center walls), after which participants were prompted to navigate to the goal. Experiments differed in terms of i) the mazes participants navigated, ii) whether the obstacles were presented before or during the execution of the plan, and iii) what the participant reported.

We reanalyzed the data of Ho and colleagues’ experiments 1 and 2 for the present study. In experiment 1 (i.e., dataset Ho 1), participants were presented with the obstacles throughout the trial. At the end of each trial, participants were asked to rate “How aware of the highlighted obstacle were you at any point?” using a nine-point scale. In experiment 2 (i.e., dataset Ho 2), participants were similarly asked to rate their awareness of the various obstacles but were required to plan their solution before they began to solve the maze. See ¹⁸ for details concerning the experimental procedures.

We did not reanalyze the results of the fourth experiment by Ho and colleagues. In this experiment, participants were not presented with all the information (i.e., obstacles) at once to solve the maze. Instead, they discovered obstacles by hovering over them with a cursor.

To further test the effects of attention on task representations, we designed a novel set of maze stimuli. This consisted of 12 mazes with task-relevant obstacles lateralized to a hemifield (left or right) and 12 non-lateralized stimuli. Each maze consisted of six obstacles, three on each hemifield, none of which crossed the veridical meridian. This ensured that there were an equal number of obstacles for computing the lateralization index (see below). Maze stimuli of both sets were equated on several nuisance covariates (see Supplemental Table S1). For dataset dSC 1, participants solved each of these 24 mazes four times (i.e., all possible orientations of the maze).

The design of the in-person experiment (i.e., dataset dSC 1) closely followed the second experiment of Ho and colleagues¹⁸. On every trial, participants were presented with a maze stimulus for 6 seconds, over which they were required to plan. The maze stimulus was offset, and participants were required to solve the maze after a one-second delay. On every trial, participants reported on their task representations using a nine-point awareness scale.

Participants

For datasets Ho 1 & 2, participants completed the task online on Prolific. In dataset Ho 1, 194 participants completed submissions, 161 of whom were included in the final sample after exclusions. In dataset Ho 2, 188 participants completed submissions, 162 of whom were included.

For dataset dSC 1, 35 participants (mean age = 23.14, SD = 5.35; 12 male) completed an in-person eye-tracking experiment (see 'Eye-tracking acquisition'). None of the participants were excluded from the data analysis. We excluded trials where participants’ reaction times were longer than 20 seconds, or where participants deviated more than nine moves from the optimal path (which reflected 3SD above the mean).

Ethics

All procedures were approved by the University College London ethics committee and adhered to the Declaration of Helsinki. Informed consent was obtained from each participant prior to each experiment.

VGC model

We fit the previously described VGC model to our maze stimuli¹⁸. Briefly, this model computes the optimal simplified task representation such that it maximizes the utility of the representation while also minimizing the cognitive cost of keeping information in mind. This model assumes that a decision-maker combines a subset of cause-effect relationships to represent their environment in aid of planning. For every possible construal, the model computes the value of a representation:

where U(π_c) is the utility of a construed plan π_c, and C(c) represents the cost of keeping that information in mind.

A task representation is selected according to a SoftMax decision rule. We then compute a marginalized probability for each obstacle being included within a construal,

where φ_Obstaclei is the cause-effect relationship for obstacle_i, P(c) is the probability that the task representation is selected, and ‖X‖ is a statement which evaluates to 1 if X is true, and 0 when X is false. We use the values of P(Obstacle_i) for every obstacle in a maze to predict participants’ awareness reports. See ¹⁸ for a detailed explanation of the computational model.

We focused our analyses on the static version of the VGC model (i.e., sVGC), whereby task representations are assumed to remain stable across planning. Our choice was informed by the design of the experiment where participants were required to plan over all obstacles at once.

Spatial proximity effects

To examine how the spatial context of information influences participants’ awareness reports, we ran a hierarchical linear regression model. First, for every obstacle in every maze, we rank-ordered all other obstacles based on spatial proximity. That is, the participant’s awareness report of the closest item to obstaclei on the trial was used as a predictor of the participant’s report of obstaclei in a hierarchical linear regression model. This yielded a regression model with 6 regression coefficients predicting participants’ awareness reports based on spatial proximity:

where (1| MazeID) and (1| ParticipantID) are random intercepts of each maze and participant, respectively, and β₁ reflects the contribution of the closest obstacle to obstaclei. We interpret any significant effects in this model as the influence of neighbouring stimuli on participants’ representations. We also fit the above hierarchical linear regression model for each participant separately. We report these individual beta coefficients in Figure 1b.

To ensure that the above spatial proximity effects were not driven by the VGC model predictions, we regressed out the effects of VCG model predictions from participants’ awareness reports, and used the residuals of the model as the dependent variable in a second regression where we similarly predicted the effects of neighbouring stimuli on representations.

We verified that these effects were not explained by the spatial smoothness of our data by conducting 1000 spatial null permutations. For every iteration, we permuted the mapping between each obstacle in a maze and their spatial location maintaining the number of neighbouring obstacles for every trial. We fit a hierarchical linear regression model using this permuted data and built a distribution of null beta coefficients to compare to our observed effects.

The sparsity of task representations

We sought to test the relationship between i) interindividual differences in attention effects and ii) the sparsity of task representations. First, we estimated the magnitude of each person’s attention effect by fitting a linear slope to the beta coefficients obtained (see Spatial proximity effects). A participant with a large negative slope, therefore, showed a larger effect of neighbouring obstacles on their representation. Second, we operationalized the sparsity of participants’ simplified representation as the variance of their awareness reports. A participant with a sparse representation shows a high variance in their awareness of different obstacles in a given maze. Last, we tested the linear monotonic relationship between the sparsity of participants’ representations and the attention effects using Spearman correlation.

Lateralization index

To test the effects of lateralization of task-relevant stimuli on participants’ awareness reports, we developed a lateralized index of task-relevance inspired by the alpha-power attention literature^35,42,43. We divided each maze into a right and left hemifield and computed the ratio of task-relevant obstacles on both sides:

where sVGC is the model’s prediction of each obstacle task-relevance for that maze. Note obstacles only with a majority of its blocks within a single hemifield were considered (3 or more squares). This yielded an index of task-relevance lateralization for each maze stimulus. We repeated the above procedure to obtain an index of task-relevance lateralization for the horizontal meridian (superior vs inferior hemifield).

We tested whether the lateralization index moderated the relationship between the value-guided model predictions (sVGC) and participants’ awareness reports using a hierarchical linear regression model.

where β₃ represents the interaction between the VGC model predictions and the lateralization index.

Spotlight-VGC model

Inspired by previous literature comparing visuospatial attention to a spotlight that moves across the visual field, we developed an extension of the VGC model to account for the effects of attentional selection in forming task representations.

To do this, we recomputed the P(Obstacle_i) as a weighted average of its neighbours. We computed the distance between every obstacle in the maze, and searched for obstacles with neighbours within 3 squares (Manhattan distance) away from obstaclei. The distance of 3 squares reflects the ‘width’ of the attentional spotlight, and was chosen based on the median distance between neighbouring obstacles from our ranked analyses (Figure 1b & Figure S2). Neighbouring obstacles that fell within the attention spotlight were averaged as follows:

where n is the number of obstacles that fall within the width of the attentional spotlight (i.e., neighbouring items). We repeat this procedure for all obstacles within each maze. If an obstacle did not have any neighbours, then the value of P(Obstacle_i) remained identical to the value of original VGC model.

We used the outputs of the attention spotlight model in a hierarchical linear regression to predict participants’ awareness reports, where we included participant and maze random intercepts:

All linear regression models were fit with the lmer package in R.

Null permutations

To ensure that the improved model fit of the attentional spotlight model was not driven by the spatial smoothness of our data, we conducted a series of control analyses where we permuted the model predictions within mazes.

To do so, we re-assigned the P(Obstaclei) of each obstacle in a given maze to a random item such that each obstacle was given a new model prediction. This permutation procedure maintains the distribution of P(Obstaclei) across obstacles for each maze, while randomizing the location of task-relevant information. We repeated this procedure for each maze separately. We then used these random model predictions to predict participants’ reports using the same hierarchical linear model described in Spotlight-VGC model. We repeated this procedure 1000 times to generate a null distribution of beta coefficients. We compared the observed beta value for the spotlight-VGC model against this distribution. We note that averaging neighbouring obstacles before or after the permutation of the model predictions qualitatively yielded the same result.

Eye-tracking acquisition

For dataset dSC1, participants completed the computer task while their eye-position and pupil size were monitored using an EyeLink 1000 Plus eye tracker at 1000Hz (SR Research, Osgoode, ON). Participants were seated comfortably in a dimly lit room in front of a 24-inch monitor set to the resolution of 1,920 x 1,080 pixels at 60 Hz. Participants were positioned 60 centimetres away from the screen and rested their heads on a mount. Stimuli were presented on MATLAB 2019a using psychtoolbox (3.0.16), synchronized with the eye tracker. Before the start of the experiment, participants completed a standard 5-point calibration procedure. Drift correction was applied after every block.

Eye-tracking preprocessing & analysis

Eye-tracking data were preprocessed with the PuPL toolbox in MATLAB⁷⁰. Impossible data points (i.e., gaze outside the screen’s bounds) were removed, in addition to data 50ms before and 150 ms after eye blinks (identified by pupillometry noise⁷¹). Segments of missing gaze position, up to 400ms long, were interpolated using cubic splines. We analyzed eye position data between −1000 ms and 6000 ms around the presentation of the maze, which corresponds to the planning window and the one second prior to planning. To verify that participants did not move their eyes more frequently during planning for lateralized mazes, we computed the standard deviation of eye position along the X-axis for each trial. We compared the fluctuations across lateralized and non-lateralized trials with a two-sample t-test. To verify the robustness of our behavioural effects, we identified and removed from further analysis trials where participants’ eye position exceeded two squares away from fixation.

Data & code availability

All in-house code used for data analysis and visualization is available on GitHub https://github.com/jasondsc/ConsciousDetour. The reanalyzed data presented herein are available from https://www.nature.com/articles/s41586-022-04743-9. The data from experiment 3 are available from https://osf.io/sa6vf/.

Acknowledgements

J.d.S.C. is supported by the Natural Sciences and Engineering Research Council of Canada (NSERC) postdoctoral fellowship program. S.M.F. is a CIFAR Fellow in the Brain, Mind and Consciousness Program and is supported by a Wellcome/Royal Society Sir Henry Dale Fellowship [206648/Z/17/Z] and UKRI under the UK government’s Horizon Europe funding guarantee (selected as ERC Consolidator, grant number 101043666).

Supplemental materials

Figure S1.

Figure S2.

Figure S3.

Figure S4.

Figure S5.

Figure S6.

Table S1.

Spatial proximity effects

Table S2.

Table S3.

Fig. S7.

Fig. S8.

Table S4.

Table S5.

We explored inter-individual differences in spatial attention effects for the second experiment where participants were required to plan their route upfront.

Similar to the first experiment, obstacles closest to the probed item positively related to participants’ awareness (1^st rank: t(161)= 19.53, p_fdr < 0.001; 2^nd rank: t(161)= 13.69, p_fdr < 0.001). In contrast, obstacles furthest from the probed item were negatively related to participants’ awareness of the probed item (4^th rank: t(161)= −5.36, p_fdr < 0.001; 5^th rank: t(161)= −7.32, p_fdr < 0.001; 6^th rank: t(156)= −10.16, p_fdr < 0.001; see Figure S9).

Fig. S9.

Fig. S10.

Table S6.

Table S7.

Fig. S11.

Task-relevant lateralization effects

Given the observed spatial attention effects and previous attention literature^39–41 we hypothesized that participants would select task-relevant information with greater ease and therefore form a representation of the stimulus that is more aligned with the predictions of the VGC model when task-relevant obstacles are presented to a single hemifield. To test this, we computed a lateralization index per maze, and used this as a moderator in a hierarchical linear regression model. A summary of the moderation effects can be found in Table S6 for both horizontal and vertical lateralization effects across both experiments.

Table S8.

Table S9.

Table S10.

Attentional spotlight model

To ensure that our attentional spotlight model results were robust, we conducted control analyses where we ran spatial permutations of our model to ensure that the observed results were not simply due to the spatial smoothness.

Fig. S12.

Sensitivity analyses

Fig. S13.

Fig. S14.

Table S11.

Table S12.

Table S13.

Table S14.

Table S15.

Table S16.

Additional information

Funding

Natural Sciences and Engineering Research Council of Canada

• Jason da Silva Castanheira

Wellcome Trust

https://doi.org/10.35802/206648

• Stephen M Fleming

Royal Society (206648/Z/17/Z)

• Stephen M Fleming

European Research Council (101043666)

• Stephen M Fleming