Feature misbinding underlying serial-order effects of visuospatial working memory

Abstract

The accurate processing of incoming visual information in serial order is fundamental to visual cognition. Prior studies have demonstrated primacy and recency effects in tasks requiring the serial recall of visual stimuli such as letters, digits, words, or locations. However, there is ongoing debate about whether these primacy/recency effects on working memory retrieval reflect variable representational precision for multiple items in memory or misbinding of item features (e.g., location) and their ordinal position. This study sought to examine potential sources contributing to serial-position effects in visuospatial working memory using eye tracking and statistical modeling approaches. In two eye-tracking experiments, we measured the latency and endpoint error of serial-order memory-guided saccades under varying cue conditions (order cue vs. quadrant cue) from a total of 92 participants. The first (primary) memory-guided saccade showed primacy and recency effects on endpoint error, latency, and transposition error in the order-cue condition, but not in the quadrant-cue condition. Probabilistic modeling of memory-guided saccade distribution showed a better fit of a standard (non-swapping) model to the quadrant-cue condition and a swap model to the order-cue condition. These findings indicate that visuospatial working memory representation varies across serial positions primarily due to location–serial–position misbinding rather than variable memory precision about location.

Introduction

The ability to track the order of incoming sensory events is critical for many daily functions, ranging from verbal to visuospatial tasks. Although serial-order verbal recall has been more extensively studied (see review by Hurlstone, Hitch, & Baddeley, 2014), relatively few researched the visuospatial domain (Bonanni, Pasqualetti, Caltagirone, & Carlesimo, 2007; Farrand, Parmentier, & Jones, 2001; Guérard & Tremblay, 2008; McAteer, Ablott, McGregor, & Smith, 2023). It is unclear how exactly multiple items in a visuospatial sequence are represented in working memory and what latent cognitive sources contribute to variable serial-position memory effects. Using eye-tracking and statistical modeling, this study sought to address these issues, examining the nature of serial-order processing in visuospatial working memory.

Items in different serial positions are not remembered equally. A plethora of verbal short-term memory studies have demonstrated primacy and recency effects on response accuracy (e.g., Guérard & Saint-Aubin, 2012) and latency (e.g., Haberlandt, Lawrence, Krohn, Bower, & Thomas, 2005) during serial recall of visual/auditory letters, digits, or words. Similar serial-position effects have been observed in visuospatial tasks, including serial reconstruction (Avons, 2007; Cortis, Dent, Kennett, & Ward, 2015; Hurlstone & Hitch, 2015; Jones, Farrand, Stuart, & Morris, 1995; Martin, Tapper, Gonzalez, Leclerc, & Niechwiej-Szwedo, 2017), serial recall (Bonanni et al., 2007; Farrand et al., 2001; Guérard & Tremblay, 2008), and partial serial recall of spatial locations, orientations, and motions (Gorgoraptis, Catalao, Bays, & Husain, 2011; McAteer et al., 2023; Udale, Tran, Manohar, & Husain, 2022; Zokaei, Gorgoraptis, Bahrami, Bays, & Husain, 2011). For instance, the accuracy of recalling spatial locations in a sequence tends to decrease from early to intermediate items and then increase for last serial positions (e.g., Guérard & Tremblay, 2008). Inter-response latency, defined as the time between successive responses, decreases sharply between the first and second serial positions and remains relatively stable thereafter (e.g., Hurlstone & Hitch, 2015).

Despite these consistent observations, the neurocognitive mechanisms underlying serial-position effects are not well understood. Features such as spatial location may be represented with varying precision across serial positions due to dynamic distribution of working memory resources among items (Gorgoraptis et al., 2011; Udale et al., 2022), as proposed in recent cognitive models of working memory such as the resource model (Bays, Catalao, & Husain, 2009; Bays & Husain, 2008). At the neural level, lower spatial working memory precision has been associated with greater variability in prefrontal neuronal activity during the delay period (Wimmer, Nykamp, Constantinidis, & Compte, 2014), suggesting a possible neural basis of variable spatial working memory precision across experimental factors such as serial position.

Additionally, serial-position effects may arise from the misbinding between visuospatial features and the serial position of different objects in a sequence. Cognitive models of serial-order processing, such as the position-marking model (Brown, Preece, & Hulme, 2000) propose that items in a sequence are bound with contextual cues such as serial positions. Similarly, resource model of working memory (e.g., Bays et al., 2009; Manohar, Zokaei, Fallon, Vogels, & Husain, 2019) suggest that misbinding among multiple features of memorized items is a major source of recall error. Such feature–serial–position binding may be achieved neurally by joint coding of spatial/visual and ordinal information (Ninokura, Mushiake, & Tanji, 2003, Ninokura, Mushiake, & Tanji, 2004) and via multiplicative gain modulation (Botvinick & Watanabe, 2007) in the lateral prefrontal cortex.

It remains unclear based on previous studies whether serial-position effects in visuospatial working memory are driven primarily by changes in memory precision, feature misbinding, or both. Free-recall studies have shown serial-position effects for spatial location in the absence of explicit binding demands, suggesting a role for memory precision (Bonanni et al., 2007; Cortis et al., 2015). However, these effects emerged only with relatively long sequences (six or more items) and were confounded by output interference. Partial-report paradigms minimize output interference, but consistently require the maintenance and binding of multiple features for retrieval, making it difficult to isolate precision from misbinding (e.g., Gorgoraptis et al., 2011, orientation + color/order; McAteer et al., 2023, location + color; and Schneegans, McMaster, & Bays, 2023, color + shape or color/shape + order). Mixture modeling has been applied to separate these sources of error (Bays et al., 2009), but findings remain mixed, with studies reporting serial-position effects on misbinding probability (Zokaei et al., 2011), memory precision (McAteer et al., 2023), or both (Gorgoraptis et al., 2011).

The present study aimed to dissociate two sources of serial-position effects in visuospatial working memory: spatial working memory precision and location–serial–position misbinding. Across two eye-tracking experiments, 92 participants completed a serial-order, memory-guided saccade (sMGS) task in which four sequentially presented locations were maintained over a variable delay, allowed by a cue indicating which of the four locations is the target to report. The quadrant-cue condition indicating the target's quadrant required participants to maintain only the target location, with no explicit demand to recall its serial position. In contrast, the order-cue condition indicating the target's serial position required participants to maintain both the location and the serial position of each item, thereby engaging both memory precision and feature-binding processes. By comparing saccadic performance across the two cue conditions and applying mixture models to trial-by-trial saccade distributions, we sought to dissociate serial-position effects arising from changes in spatial memory precision versus location–serial–position misbinding.

We hypothesized that serial-position effects are driven by both location–serial–position misbinding and changes in spatial working memory precision across memorized items. Specifically, we predicted that, in the quadrant-cue condition (implicit serial-position demand), recall error, latency, and memory precision would vary across serial positions, reflecting changes in spatial memory precision. The order-cue condition (explicit serial-position demand) was expected to show stronger serial-position effects on recall error, latency, and swap rates compared with the quadrant-cue condition, reflecting the additional demands of binding location to serial position.

Experiment 1

Experiment 1 examined the effects of serial position on visuospatial working memory under the quadrant-cue and order-cue conditions, manipulated within participants. Eye gaze positions were recorded from 65 participants while performing the sMGS task with the order-cue and quadrant-cue conditions as a within-subject factor. Error and latency of the primary and closest saccades to the target were analyzed and modeled across serial positions and cues.

Participants

We recruited 65 college students from the Department of Psychology's participant pool (mean, 20.4 ± 3.6 years; 39 females). All participants had normal or corrected-to-normal vision and no self-reported neurological or psychiatric conditions. Informed consent was obtained from all participants. All the experimental procedures followed the protocol approved by the Institutional Review Board committee at Stony Brook University. Course credits were given to the participants upon completion of the study session.

Visual stimuli

Visual stimuli used in the working memory task included eight images: two female faces, two male faces, and four houses. The face stimuli had neutral expressions and were selected from the NimStim database (Tottenham et al., 2009). The house stimuli were selected from the DalHouse database (Filliter, Glover, McMullen, Salmon, & Johnson, 2016). All stimuli were processed to yield the same resolution (500 × 500 pixels), mean, and standard deviation of luminance (CIE Lab; mean, 48.7± 24.3) using custom scripts in MATLAB and manual editing in Adobe Photoshop. The size of all image stimuli was 0.8 degrees of visual angle (dva), and the size of the center cross and cues was 0.5 dva in diameter.

Visual stimuli were presented in 32 prototype locations (Fig. 1C). The eccentricity of the images was either 3.0 or 5.5 dva. The polar angle of the images was 15°, 35°, 55°, or 75° within each quadrant. This uneven spacing of spatial locations was designed to minimize the use of verbal strategies, repulsion biases, and perceptual ambiguity near the cardinal axes. Target locations were jittered across trials to reduce potential long-term memory effects, with the polar angle jittered between −5° and 5° and eccentricity jittered between −0.5 and 0.5 dva using a uniform distribution around the target.

Figure 1.

sMGS task. (A) A schematic diagram of key events in a trial. Four items (a cartoon illustration of the stimuli is shown in the figure) were sequentially displayed in different locations on the screen with a 0.5-second interstimulus interval (not shown). After a delay of either 4, 6, 8, or 10 seconds, a task cue was shown to indicate the response target. Accordingly, the participants shift their eye gaze to the remembered target location. The gray curve illustrates a possible eye movement path during the response window (1 second), which was not presented in the actual task. (B) The two types of cues used in the experiment. During the quadrant-cue condition, one-quarter of the fixation stimulus turns black, indicating the response target's quadrant. During the order-cue condition, a number appears at the center of the fixation stimulus, indicating the response target's serial position. (C) Target locations and sample sequence structures. There are 32 possible prototype locations, including 16 different polar angles (range between 15° and 345°) and two different eccentricities (3.0 and 5.5 dva). On each trial, four items, one per quadrant at the same eccentricity, are presented sequentially either from left to right or from right to left to form four possible spatiotemporal patterns. Example spatial locations and the temporal order of the four items in a trial are displayed. (D) Sample gaze time course from all trials of a representative participant. Preprocessed eye gaze position in dva is plotted from fixation onset to the end of the response period. Horizontal (x) eye positions are plotted in red, and vertical (y) eye positions are plotted in blue.

Tasks were programmed using Experimental Builder (SR Research, Ontario, Canada, Version 2.3.38) and displayed using a PC with Windows 10 operating system. The tasks were delivered on a Dell monitor with a resolution of 1,280 × 1,024 and a refresh rate of 75 Hz. Participants viewed the screen from a distance of 67.7 cm. The screen's height was 30.5 cm, and its width was 38.0 cm, resulting in a maximum field of view of 31.4 dva horizontally and 25.4 dva vertically. Gaze positions were recorded monocularly at a sampling rate of 500 or 1,000 Hz using a desktop-mounted, video-based eye-tracker, the Eyelink 1000 (SR Research).

Task design and procedure

All participants completed a 1.5-hour study session consisting of eight runs of the sMGS task. Each trial of the task (Fig. 1A) began with a 1.5-second fixation period, followed by the sequential presentation of four visual stimuli in different locations for 500 ms with a 500-ms interstimulus interval and they appeared at the same eccentricity (3.0 or 5.5 dva), one in each quadrant. Following a delay of either 4, 6, 8, or 10 seconds, a cue (quadrant or order) appeared in the center of the screen, indicating which of the four remembered items is the response target (Fig. 1B). Specifically, the quadrant cue indicated the target quadrant by turning one-quarter of the fixation circle black, whereas the order cue indicated the serial position of the response target with a number displayed at the center of the fixation circle. Participants were instructed to shift their gaze to the remembered target location as quickly and accurately as possible within a 1.5-second response window. After completing the saccade(s), participants returned their gaze to the central fixation cross and maintained fixation until the next trial began. The intertrial interval varied between 1.25, 1.50, and 1.75 seconds, with an average of 1.50 seconds.

The main independent variables in this study were cue type and target serial position, which were manipulated within participants. Each participant completed four runs (64 trials) of the order cue task and four runs (64 trials) of the quadrant cue task, with the order of two tasks counterbalanced between participants. We probed all four serial positions for each cue condition an equal number of times.

In addition to the cue type and serial position, we controlled the spatiotemporal pattern of the sequence. The four items were presented in order from either the left to the right hemifield or from the right to the left hemifield in four temporal patterns (Fig. 1C). They appeared in 54 spatial forms (Supplementary Fig. S1) at two levels of eccentricities (3.0 dva and 5.5 dva). The spatial forms of the sequences were carefully selected to be either fully asymmetric or have only one pair of dots symmetric relative to the x or y axis or the origin. These spatiotemporal patterns were designed to prevent participants from predicting target location using sequence structure and to reduce both novelty and long-term memory effects.

The same spatiotemporal patterns were used for both cue conditions. Each of the four sequence patterns appeared four times in each run, and we probed the four serial positions for each of the four temporal patterns equally, resulting in 16 trials in total. Within a run, spatial forms were pseudo-randomly distributed across trials while meeting the following criteria: 1) spatial patterns were maximally distinct across trials without repetition, based on a Procrustes analysis; 2) all 16 locations from one eccentricity appeared the same number of times; and 3) all 4 quadrants were probed the same number of times. Each run included targets of a fixed eccentricity (e.g., 3 dva) and different eccentricities were counterbalanced between runs.

Last, we manipulated the order of the visual images (face first vs. house first) within participants and the delay duration between participants (4 and 6 seconds, 30 participants; 8 and 10 seconds, 35 participants). The order of the images was counterbalanced between runs orthogonal to the eccentricities. After counterbalancing, we pseudo-randomly shuffled the trials to ensure the sequence was as unpredictable as possible.

Oculomotor analysis

Eye gaze data were preprocessed using customized MATLAB scripts. The raw data were smoothed using a second-order Savitzky–Golay filter with a kernel size of 20 ms (Nyström & Holmqvist, 2010). Artifacts such as blinks and noise were removed. Blinks were identified as samples with pupil velocity exceeding 8,000 pixels per second, and an additional 100 ms of data before and after each blink were excluded to clean up surrounding noise. Samples with gaze positions outside the screen, gaze velocity greater than 1,000 dva/s, or acceleration greater than 100,000 dva/s² (Nyström & Holmqvist, 2010) were identified as noise and removed without interpolation.

Following blink and artifact removal, an offline drift correction was performed (Vadillo, Street, Beesley, & Shanks, 2015). Baseline gaze positions were identified for each trial as fixations occurring 750 ms before stimulus onset. Fixations were defined as samples with a velocity of 10 dva/s or less and a duration of 30 ms or greater, and with both horizontal and vertical gaze dispersion of 1.5 dva or less. For each trial, a linear transformation matrix was estimated as the mean distance between the mean baseline gaze location and the fixation cross location. This transformation matrix was applied to the horizontal and vertical gaze positions of all samples within that trial.

Saccades in the preprocessed data were classified as samples with a gaze velocity of 30 dva/s or greater, acceleration of 6000 dva/s² or greater, amplitude of 0.25 dva or greater, and duration of 8 ms or greater. We defined several types of saccades for analysis. The first (primary) saccade within the 1,500-ms response window after cue onset was analyzed to estimate participants’ initial memory recall on a trial-by-trial basis. Additionally, within the same time window, we identified the closest saccade to the target within 3.5 dva of the target and at least 2.0 dva from the fixation cross to estimate potential corrective saccades made to the target.

The endpoint error for both the primary and closest saccades was calculated as the Euclidean distance between the mean position of the first fixation after the saccade and the target position in degrees of visual angle. The latency of the primary and closest saccades was measured as the time between cue onset and saccade onset in milliseconds.

Transposition errors were estimated by the percentage of trials in which participants responded to a non-cued stimulus. We calculated the polar angle of the primary saccade endpoint and compared it with the polar angle of the target. If the primary saccade landed within a different quadrant from the target, that trial was defined as having a transposition error.

All participants were included in the final analysis. We excluded the following types of trials from the final analysis: 1) there was a blink or an artifact during the first 500 ms of the response window (9.0% ± 9.9% of trials on average across participants); 2) there was no measurable baseline (0.8% ± 1.1% of trials); 3) there was no saccade detected during the response window (4.7% ± 6.4% of trials); 4) the primary saccade latency was less than 50 ms (0.08% ± 0.34% of trials); 5) the primary saccade start position was larger than 2 dva from the center cross (2.8% ± 3.5% of trials); 6) there was no fixation detected after the primary saccade (1.6% ± 3.4% of trials); or 7) other data collection errors, such as missing pupil (1.0% ± 4.6% of trials).

After trial exclusion, on average 80.0% ± 14.0% of trials were included in the final analysis across participants. Among these trials, 65.7% ± 15.6% of trials were identified as responses to the target as their primary saccade endpoint is less than 3.5 dva from the target location, 5.0% ± 3.2% of trials as responses to non-target (a saccade endpoint that is more than 3.5 dva from the target but less than 3.5 dva from one or more non-targets), 9.4% ± 6.7% of trials as responses outside of all regions of interest (a saccade endpoint more than 3.5 dva from any target or non-target). The percentage of trials included for the quadrant-cue condition (80.1% ± 14.5%) did not significantly differ from the order-cue condition, 80.0% ± 16.1%; t(64) = 0.06, p = 0.95.

Model design and fitting

To dissociate latent sources of error in serial-order spatial working memory, we applied two-dimensional probability mixture models (MemFit2D; Grogan et al., 2020) to primary saccade endpoints across trials. The endpoint distribution was modeled by a weighted mixture of Gaussian distributions centered on the target and non-target locations and a uniform distribution representing random guessing. We modeled primary saccades rather than closest gaze positions to the target to capture misbinding errors and random guesses that would otherwise be obscured by post hoc endpoint selection.

We compared three models of increasing complexity. The standard mixture model (Zhang & Luck, 2008) estimates target-directed responses and random guessing. The probability of responses, p(x), is modeled by a weighted mixture of a bivariate Gaussian centered on the target location Τ with variance σ² (σ_x² = σ_y² = σ², σ_xy² = 0), and a bivariate uniform distribution spanning the area of the entire screen, A, capturing responses not systematically related to any stimulus location. These distributions were weighted by α and γ, respectively, where α + γ = 1 and 0 ≤ α, γ ≤ 1:

$$\begin{array}{c}\hfill p\left(x\right)=\alpha N\left(x-\mathbf{T}\left|{\sigma }^2\right.\right)+\gamma \frac{1}{A}.\end{array}$$

To further model item swapping, where participants bind the correct item location with the wrong serial position, we applied the swap model (Bays et al., 2009). Besides target-related responses and guessing, the swap model estimates an additional response directed toward non-target locations by three bivariate Gaussian distributions (m = 3), each centered around a non-target locations NT_i sharing the same mean and variance as the target distribution. The target-related, non–target-related, and guessing response distributions were weighted by α, β/m, and γ respectively, where α + β + γ = 1 and 0 ≤ α, β, γ ≤ 1:

$$\begin{array}{ccc}& & p\left(x\right)=\alpha N\left(\mathrm{x}-\mathbf{T}\left|{\sigma }^2\right.\right)\hfill \\ & & +\beta \frac{1}{\mathrm{m}}\sum _{i=1}^{m}N\left(\mathrm{x}-\mathbf{N}{T}_i\left|{\sigma }^2\right.\right)+\gamma \frac{1}{A}.\hfill \end{array}$$

Last, we applied the swap plus systematic bias model (Grogan et al., 2020) to capture the systematic bias in responses independent of memory precision. On top of the swap model, we allowed the mean of the bivariate target distribution to vary (µ_x = µ_y = µ) to dissociate biases (indexed by µ) and variances of responses (indexed by σ²):

$$\begin{array}{ccc}& & p\left(x\right)=\alpha N\left(\mathrm{x}-\mathbf{T}\left|\mu ,{\sigma }^2\right.\right)\hfill \\ & & +\beta \frac{1}{\mathrm{m}}\sum _{i=1}^{m}N\left(\mathrm{x}-\mathbf{N}{T}_i\left|\mu ,{\sigma }^2\right.\right)+\gamma \frac{1}{A}.\hfill \end{array}$$

We fitted the models to the data using maximum likelihood estimation and derived the following parameters: α, probability of responding to target; β, probability of responding to non-target (reflecting item swapping or feature misbinding); γ, probability of random guessing; µ, systematic bias of responses; and σ, memory imprecision (reciprocal of memory precision).

Model fitting was conducted on 13.3 trials on average (range, 3–16 trials) per serial position per cue type for each participant. The Akaike information criterion (AIC) (Akaike, 1974) and Bayesian information criterion (BIC) (Schwarz, 1978) were used to assess the goodness of fitting. Lower AIC or BIC values indicate better-fitting models. A difference in AIC or BIC values of greater than 2 between models provides substantial evidence favoring the model with the lower value (Burnham & Anderson, 2004; Raftery, 1995).

Statistical analyses

Two-way repeated measures analyses of variance (ANOVAs) with serial position and cue type as within-subject factors were performed for the following dependent variables: endpoint error and latency of the primary saccades and the closest saccades to the target; transposition error of the primary saccades; and model-derived parameters of the primary saccades, including target rate (α), swap rate (β), guess rate (γ), and memory imprecision (σ). Additional nonparametric aligned rank transform ANOVAs (ART-ANOVA; Wobbrock, Findlater, Gergle, & Higgins, 2011) were conducted using the ARTool package in R (Kay, Elkin, & Wobbrock, 2021) on all four model parameters due to their significant violation of normality assumptions.

Results

Memory-guided saccade error across serial positions

To examine how endpoint errors of memory-guided saccade vary by serial positions across cues, we conducted two-way repeated measures ANOVAs on the primary and closest saccades to the response target with serial position and cue type as within-subject factors (Figs. 2A, 2B).

Figure 2.

Memory-guided saccade endpoint errors across cue types and serial positions in Experiment 1. Plots in the top row show saccade endpoint error for the two cue conditions at each serial position for the primary saccade (A) and the saccade that was closest to the target within the area of interest (B). Plots in the bottom row show the corresponding saccade latency (C, D). The y axes represent the saccade measures and the x axes represent the serial position of the target in the stimulus sequence. Quadrant cue is plotted in gray and rectangle symbol; order cue is plotted in black and circle symbol.

For primary saccadic responses (Fig. 2A), we observed significant effects of cue, F(1, 64) = 60.40, p < 0.001, ${\eta }_{\mathrm{p}}^2$ = 0.49; serial position, F(3, 192) = 9.36, p < 0.001, ${\eta }_{\mathrm{p}}^2$ = 0.13; and their interaction, F(3, 192) = 3.05, p = 0.03, ${\eta }_{\mathrm{p}}^2$ = 0.05. The interaction seemed to be driven by a significant main effect of serial position for the order-cue condition, F (3, 192) = 7.56, p < 0.001, ${\eta }_{\mathrm{p}}^2$ = 0.11, but not the quadrant-cue condition, F (3, 192) = 2.25, p = 0.08, ${\eta }_{\mathrm{p}}^2$ = 0.03. Polynomial contrast analysis of the order-cue condition revealed a significant quadratic contrast, t(192) = −4.32, p < 0.001, but not linear contrast, t(192) = −1.28, p = 0.20, in saccade endpoint errors across serial positions. Pairwise comparisons of serial positions revealed significant differences between positions one and two, t(64) = −3.63, p = 0.002, Cohen's d = −0.37, and between positions two and four, t(64) = 4.35, p < 0.001, Cohen's d = 0.45, in the order-cue condition.

In contrast, the endpoint error of the closest saccades toward the target (Fig. 2B) was significantly modulated by serial position, F(3, 192) = 4.81, p = 0.003, ${\eta }_{\mathrm{p}}^2$ = 0.07, but not by cue type, F(1, 64) = 0.29, p = 0.59, ${\eta }_{\mathrm{p}}^2$ = 0.004, and there was no significant interaction between them, F(3, 192) = 0.61, p = 0.61, ${\eta }_{\mathrm{p}}^2$ = 0.009. Simple main effects of serial position were significant for the order cue, F(3, 192) = 3.68, p = 0.013, ${\eta }_{\mathrm{p}}^2$ = 0.05, but not the quadrant-cue condition, F (3, 192) = 1.72, p = 0.16, ${\eta }_{\mathrm{p}}^2$ = 0.03. In the order-cue condition, the linear contrast across serial positions were significant, t(192) = −2.91, p = 0.004, but the quadratic contrast was not, t(192) = −1.21, p = 0.23; the endpoint error differed significantly between serial positions two and four, t(64) = 3.03, p = 0.017, Cohen's d = 0.46.

Together, these results showed both primacy and recency effects on the primary memory-guided saccades and a recency effect on the closest saccade to the target during the order-cue condition. Memory-guided saccade errors during the quadrant-cue condition, however, did not significantly vary by serial position.

Memory-guided saccade latency across serial positions

To examine how latency of memory-guided saccade varies by serial positions across cue conditions, we conducted similar two-way repeated measures ANOVAs on saccade latency (Figs. 2C, 2D). For primary saccades (Fig. 2C), significant effects were found for serial position, F(2.6, 166.6) = 20.78, p < 0.001, ${\eta }_{\mathrm{p}}^2$ = 0.25; cue type, F(1, 64) = 217.78, p < 0.001, ${\eta }_{\mathrm{p}}^2$ = 0.77; and their interaction, F(3, 192) = 25.75, p < 0.001, ${\eta }_{\mathrm{p}}^2$ = 0.29. The interaction was driven by the serial-position effect in the order-cue condition, F (3, 192) = 38.91, p < 0.001, ${\eta }_{\mathrm{p}}^2$ = 0.38; with both linear, t(192) = 6.11, p < 0.001; and quadratic contrasts, t(192) = −8.75, p < 0.001, reaching significance for saccade latency. During the order-cue condition, all pairwise comparisons between serial positions reached statistical significance (ps ≤ 0.001) except for the difference between serial positions two and four (p = 0.08). There was no significant serial-position effect in the quadrant-cue condition, F (3, 192) = 0.86, p = 0.46, ${\eta }_{\mathrm{p}}^2$ = 0.01.

Similar effects were observed for closest saccades to the response target (Fig. 2D), including serial position, F(2.7, 175.8) = 24.69, p < 0.001, ${\eta }_{\mathrm{p}}^2$ = 0.28; cue type, F(1, 64) = 175.33, p < 0.001, ${\eta }_{\mathrm{p}}^2$ = 0.73; and their interaction, F(3, 192) = 21.74, p < 0.001, ${\eta }_{\mathrm{p}}^2$ = 0.25. Latency varied significantly across serial positions, F(3, 192) = 36.77, p < 0.001, ${\eta }_{\mathrm{p}}^2$ = 0.37, in the order-cue condition, with both linear, t(192) = 6.68, p < 0.001, and quadratic, t(192) = −8.02, p < 0.001, contrasts reaching significance. During the order-cue condition, all pairwise comparisons between serial positions were significant (p < 0.005) except between positions two and four (p = 0.28). There was no significant serial-position effect in the quadrant-cue condition, F (3, 192) = 1.42, p = 0.24, ${\eta }_{\mathrm{p}}^2$ = 0.02.

Together, the latency of both the primary and closest memory-guided saccades to target showed a primacy and recency effect during the order-cue condition but no significant serial-position effects found for the quadrant-cue condition.

Modeled memory precision, swap rate, and transposition errors across serial positions

To characterize the sources of serial-position effects on memory-guided saccades, we fitted three statistical models to the endpoint distribution of primary saccades.

The standard mixture model, incorporating only target and guessing distributions, best fitted the quadrant-cue data (Table 1). The order-cue data were best fitted by the swap model with the additional non-target distributions included in the model. Model fit was evaluated using the AIC and by the standard mixture model according to the BIC (Table 1). These results indicate increased item swapping in the order-cue condition.

Table 1.

Model fit for Experiments 1 and 2. Values are mean ± standard deviation.

	Standard mixture model		Swap model		Swap model with systematic bias
Cue type	AIC	BIC	AIC	BIC	AIC	BIC
Experiment 1
Quadrant cue	300.6 ± 56.2	305.4 ± 56.6	302.6 ± 56.2	309.8 ± 56.8	304.8 ± 56.0	316.8 ± 57.0
Order cue	296.0 ± 79.7	300.6 ± 80.5	295.0 ± 79.1	302.0 ± 80.2	296.8 ± 79.1	308.4 ± 81.0
Experiment 2
Quadrant cue	894.4 ± 288.1	866.8 ± 298.7	895.2 ± 289.6	905.6 ± 290.6	896.4 ± 290.0	913.6 ± 290.7
Order cue	860.0 ± 298.0	866.8 ± 298.7	852.5 ± 294.5	862.6 ± 295.5	852.4 ± 293.8	869.2 ± 295.5

The swap model was then used to estimate parameters including memory imprecision, swap rate, guess rate, and target rate for the two cue conditions and serial positions. There was a significant effect of cue type on memory imprecision (Fig. 3A), F(1, 64) = 9.82, p = 0.003, ${\eta }_{\mathrm{p}}^2$ = 0.13. Similar to the unmodeled data described elsewhere in this article, the order-cue responses showed lower modeled memory precision (i.e., larger endpoint error) than the quadrant-cue responses, t(64) = 3.13, p = 0.003, Cohen's d = 0.27. Effects of serial position on memory imprecision approached statistical significance, F(3, 192) = 2.32, p = 0.076, ${\eta }_{\mathrm{p}}^2$ = 0.04. There was no interaction between cue type and serial position, F(3, 192) = 0.31, p = 0.82, ${\eta }_{\mathrm{p}}^2$ = 0.005. Simple main effects of serial position did not reach statistical significance during either the quadrant cue, F(3, 192) = 2.05, p = 0.11, ${\eta }_{\mathrm{p}}^2$ = 0.03, or the order-cue condition, F (3, 192) = 0.92, p = 0.43, ${\eta }_{\mathrm{p}}^2$ = 0.01.

Figure 3.

The swap model results and transposition error of memory-guided saccades in Experiment 1. (A) Modeled memory imprecision (standard error) in dva. Black and gray lines, respectively, represent the order and quadrant-cue condition. (B) Modeled swap or misbinding rate (beta). (C) Modeled target rate (alpha). (D) Modeled guess rate (gamma). (E) Number of trials (group average) during which the saccadic response was initiated toward the different serial positions. Lines of the same color represent trials of different target serial positions for the same cue condition. If the responses were always toward the correct item position, we would expect to see four clear peaks centered around each serial position as observed for the quadrant-cue condition (gray lines). During the order-cue condition, some of the saccadic responses were toward the nontargets (black lines) with lower peaks and higher valleys relative to the gray lines. (F) Transposition error, or the percentage of trials that included saccades responded to the non-target quadrant obtained from the empirical data.

For guess rate (Fig. 3D), only the effect of cue reached statistical significance, F(1,64) = 4.95, p = 0.03, ${\eta }_{\mathrm{p}}^2$ = 0.07, with a higher rate of guessing for the order cue than the quadrant cue, t(64) = 2.23, p = 0.03, Cohen's d = 0.18.

Swap rate (Fig. 3B) significantly differed by the cue, F(1, 64) = 74.6, p < 0.001, ${\eta }_{\mathrm{p}}^2$ = 0.54, and was higher in the order-cue condition compared with the quadrant-cue condition, t(64) = 8.64, p < 0.001, Cohen's d = 1.06. There was a significant serial-position effect, F(3, 192) = 7.85, p < 0.001, ${\eta }_{\mathrm{p}}^2$ = 0.11, or interaction between cue and serial position, F(3, 192) = 8.02, p < 0.001, ${\eta }_{\mathrm{p}}^2$ = 0.11. The swap rate varied significantly with serial position during the order cue, F(3, 192) = 7.95, p < 0.001, ${\eta }_{\mathrm{p}}^2$ = 0.11, with a significant quadratic contrast, t(192) = −4.76, p < 0.001, but it did not vary with serial position during the quadrant-cue condition, F (3, 192) = 0.13, p = 0.94, ${\eta }_{\mathrm{p}}^2$ = 0.002.

Consistent with the modeled swap rate, transposition errors of the primary saccades (Figs. 3E, 3F) showed a significant effect of cue, F(1, 64) = 144.92, p < 0.001, ${\eta }_{\mathrm{p}}^2$ = 0.69, serial position; F(3, 192) = 8.31, p < 0.001, ${\eta }_{\mathrm{p}}^2$ = 0.12; and their interaction, F(3, 192) =6.67, p < 0.001, ${\eta }_{\mathrm{p}}^2$ = 0.09. Particularly, transposition errors varied significantly by serial position during the order-cue condition, F(3, 192) = 0.69, p < 0.001, ${\eta }_{\mathrm{p}}^2$ = 0.13, with a significant quadrant trend, t(192) =−5.25, p < 0.001, but not during the quadrant-cue condition, F(3, 192) = 1.33, p = 0.27, ${\eta }_{\mathrm{p}}^2$ = 0.02.

As the modeled outputs significantly deviated from normality (Shapiro–Wilk tests, all ps < 0.001), we conducted additional ART-ANOVAs on all model parameters. The results were largely consistent with the standard ANOVAs reported above, confirming the Significant cue × Serial position interactions on swap rate and target rate, but not on memory imprecision or guess rate (Supplementary Table S1).

Control analyses: Delay duration, target eccentricity, and corrective saccades

We examined whether delay duration, target eccentricity, or corrective saccades may have influenced the observed serial-position effects. Delay duration (4 and 6 seconds vs. 8 and 10 seconds) did not significantly affect saccade error or interact with serial position or cue type (Supplementary Fig. S2). Target eccentricity (3.0 dva vs. 5.5 dva) affected overall saccade performance but not serial-position effects (all Eccentricity × Serial position and three-way interactions with ps > 0.28) (Supplementary Fig. S3).

When examining the role of corrective saccades on task performance, we found that corrective saccade rates showed a Serial position × Cue-type interaction, with higher rates for later positions in the order-cue condition only (Supplementary Fig. S5). However, serial-position effects on saccade performance were similar on one-step (target reached with a single saccade) and multi-step trials (requiring corrective saccades), confirming that serial-position patterns reported in the main analysis were not driven by differences in corrective saccade rates across serial positions (Supplementary Fig. S6). We further examined whether corrective saccades may compensate for the elevated item swap rate observed for the order-cue condition (Supplementary Fig. S7). Trials of correctable transpositions, with primary saccade in a non-target quadrant and a subsequent corrective saccade reaching the target, were more frequent in the order-cue than the quadrant-cue condition. However, trials of uncorrectable transpositions, with subsequent corrective saccades, also failed to reach the target and were significantly more frequent in the order cue than the quadrant-cue condition. These results suggest that the elevated swap rate observed for the order-cue condition more likely reflects a greater contribution from memory misbinding than from correctable retrieval difficulties.

Results of experiment 1

In sum, our results demonstrated a significant effect of cue type and Cue × Serial position interaction effects on memory-guided saccades. The order-cue condition, compared with the quadrant-cue condition, resulted in higher saccade endpoint error, saccade latency, transposition error, modeled swap rate, and guess rate, as well as lower modeled memory precision. The serial-position effects were evident primarily under the order-cue condition, with significant primacy and recency effects on the primary saccade endpoint error, latency, transposition error, and modeled swap rate. Additionally, a small recency effect was observed on the closest saccade to the target, and there was no significant serial-position effect on the modeled memory precision. These findings suggest that serial-position effects (both primacy and recency) are more pronounced in the multi-feature order-cue condition than in the single-feature quadrant-cue condition, and that these effects are driven more by misbinding of spatial locations to serial positions than by changes in memory precision.

Experiment 2

To replicate these findings and obtain more reliable model estimates, we conducted a second experiment that collected more trials from a separate group of participants. In this experiment, we manipulated cue conditions between participants and increased the number of trials to up to 64 per serial position per cue condition for better modeling of the saccadic responses.

Participants

Twenty-seven participants were recruited from the Psychology Department's subject pool (mean, 20.5 ± 4.1 years; 16 females) for Experiment 2. All participants had normal or corrected-to-normal vision and no self-reported neurological or psychiatric conditions. Informed consent was obtained from all participants, and the experimental procedure followed the same protocol approved by the Institutional Review Board committee at Stony Brook University as in Experiment 1.

Task design and procedure

The task design and stimuli were similar to Experiment 1, with only the following differences. First, the delay duration was reduced to 4 or 6 seconds, as a supplementary analysis of Experiment 1 showed no significant effects of delay on task performance (Supplementary Fig. S2). Second, the cue type (quadrant or order cue) was manipulated between subjects. Third, each participant completed 12 to 16 runs of the sMGS task, with 16 trials per run, over a 2.0- to 2.5-hour session. Hence, the number of trials was increased to 192 to 256 per subject. To assess performance stability across this longer experiment, we tested for effects of time (run): Although saccade error varied significantly across runs, this effect appeared to reflect improvements following rest breaks rather than progressive deterioration due to fatigue. Saccade latency did not vary with run (see Supplementary Fig. S4 for these results.) Sixteen of the 27 participants completed the quadrant-cue and 11 completed the order-cue task.

Oculomotor analysis

The oculomotor preprocessing procedure and analysis were the same as Experiment 1.

All participants completed more than 16 trials per serial position per cue. Of the 256 possible trials in total, on average 11.4% ± 9.1% of trials were excluded across participants for an artifact during the first 500 ms of the response window: 1.9% ± 2.3% of trials with no defined baseline during the fixation period, 6.5% ± 10.6% of trials with no saccade detected during the response window, 0.2% ± 0.4% of trials with a primary saccade latency shorter than 50 ms, 3.0% ± 3.7% for primary saccade starting position larger than 2 dva from the center cross, and 1.1% ± 1.2% of trials with no fixation detected after the saccade. Participants did not complete 20.2% ± 20.2% of trials on average. After excluding these trials above, 55.8% ± 21.1% of trials were included in the final analysis for the 27 participants, breaking down by 44.7% ± 20.4% of trials as responses to target, 3.7% ± 4.1% of trials as responses to non-target, and 7.4% ± 6.3% of trials as responses outside of all regions of interest.

On average, 36.5 ± 13.8 trials were obtained per serial position (range, 20.8–60.5 trials) for the quadrant cue task and 34.6 ± 13.6 trials (range, 17.3–60.3 trials) for the order-cue condition. The number of trials included does not significantly differ between cue conditions, t(10) = 0.93, p = 0.37.

Model fitting and statistical analyses

Model fitting and statistical analyses were similar to Experiment 1, except that we conducted a mixed-design ANOVA with serial position as a within-subject factor and with cue condition as a between-subject factor.

Results of experiment 2

Memory-guided saccade error across serial positions

To test for the effects of serial position, cue type, and their interaction on memory-guided saccade endpoint errors, we performed a two-way mixed-design ANOVA (Figs. 4A, 4B). The analysis showed that errors in the primary saccade significantly varied by serial position, F(3, 75) = 4.50, p = 0.006, ${\eta }_{\mathrm{p}}^2$ = 0.15, but not by cue type, F(1, 25) = 0.42, p = 0.52, ${\eta }_{\mathrm{p}}^2$ = 0.02. There was also a significant interaction between serial position and cue type, F(3, 75) = 4.21, p = 0.008, ${\eta }_{\mathrm{p}}^2$ = 0.14. A post hoc analysis revealed that the main effect of serial position was significant for the order-cue condition, F(3, 30) = 5.20, p = 0.005, ${\eta }_{\mathrm{p}}^2$ = 0.34, but not for the quadrant-cue condition, F(3, 45) = 0.20, p = 0.90, ${\eta }_{\mathrm{p}}^2$ = 0.01. In the order-cue condition, the linear contrast across serial positions was significant for saccade error, t(30) = −3.25, p = 0.003, with a significant difference between positions two and four, t(75) = 0.69, p = 0.005, Cohen's d = 1.00. The difference between positions 1 and 2 did not reach statistical significance, t(75) = −0.21, p = 0.52, Cohen's d = −0.30, although the effect size was comparable with that observed in Experiment 1 (Cohen's d = −0.37).

Figure 4.

(A–D) Saccade endpoint error and latency across cue types and serial positions in Experiment 2. Figure notations are the same as Fig. 2.

For the closest saccade to the response target, none of the mixed-design ANOVA effects reached statistical significance, although the effect of serial position approached the significance threshold, F(3, 75) = 2.69, p = 0.052, ${\eta }_{\mathrm{p}}^2$ = 0.10. A one-way repeated measures ANOVA for each cue condition revealed significant serial-position effects for the order cue, F(3, 30) = 3.26, p = 0.03, ${\eta }_{\mathrm{p}}^2$ = 0.24, but not for the quadrant-cue condition, F(3, 45) = 0.09, p = 0.97, ${\eta }_{\mathrm{p}}^2$ = 0.006. Specifically, in the order-cue condition, the linear contrast across serial positions was significant for closest saccade error, t(30) = −2.43, p = 0.02, although no post hoc pairwise comparisons between serial positions were statistically significant.

These findings align with those in Experiment 1, particularly in showing a significant serial-position effect on memory-guided saccade errors during the order-cue condition, but not during the quadrant-cue condition. However, unlike Experiment 1, no significant differences in saccade endpoint errors were observed between cue types.

Memory-guided saccade latency across serial positions

Similarly, we conducted mixed-design ANOVAs to examine the effects of serial position, cue type, and their interaction on memory-guided saccade latency (Figs. 4C, 4D). The latency of both the primary and the closest saccades to the target showed significant main effects of cue type: primary, F(1, 25) = 8.68, p = 0.007, ${\eta }_{\mathrm{p}}^2$ = 0.26; closest, F(1, 25) = 8.59, p = 0.007, ${\eta }_{\mathrm{p}}^2$ = 0.26; serial position: primary, F(3, 75) = 24.65, p < 0.001, ${\eta }_{\mathrm{p}}^2$ = 0.50; closest, F(3, 75) = 9.66, p < 0.001, ${\eta }_{\mathrm{p}}^2$ = 0.28; and their interaction: primary, F(3, 75) = 14.21, p < 0.001, ${\eta }_{\mathrm{p}}^2$ = 0.36; closest, F(3, 75) = 5.62, p = 0.002, ${\eta }_{\mathrm{p}}^2$ = 0.18. Specifically, in the order-cue condition, saccade latency varied significantly across serial positions for both the primary, F(3, 30) = 30.90, p < 0.001, ${\eta }_{\mathrm{p}}^2$ = 0.76, and closest saccades, F(3, 30) = 11.16, p < 0.001, ${\eta }_{\mathrm{p}}^2$ = 0.53. For both measures, the linear contrasts: primary, t(30) = 4.11, p < 0.001; closest, t(30) = 2.72, p = 0.01; and quadratic contrasts: primary, t(30) = −8.50, p < 0.001; closest, t(30) = −4.73, p < 0.001, were significant. No significant variations in saccade latency across serial positions were observed under the quadrant-cue condition for either type of saccade (ps > 0.1).

These results were consistent with Experiment 1, demonstrating a significant effect of cue type on memory-guided saccade latency. Serial-position effects, including both primacy and recency, were significant only during the order-cue condition and not the quadrant-cue condition.

Memory precision, swap rate, and transposition errors across serial positions

To examine the sources of serial-position effects, including memory imprecision, swapping, and guessing, we performed statistical modeling on errors of primary memory-guided saccades in Experiment 2 similar to Experiment 1. The standard mixture model best fitted the quadrant-cue data, with the lowest AIC (894.4 ± 288.1) and BIC (866.8 ± 298.7) values compared with other models. For the order cue data, the swap model had the lowest BIC (862.6 ± 295.5), while the swap plus systematic bias model had the lowest AIC (852.4 ± 293.8), indicating more item swapping in the order-cue condition (Table 1, lower half).

To differentiate memory imprecision, swap, and guess rates across cue conditions, we used the swap model to fit the primary saccade endpoint distributions for both cue conditions. Following the model estimation, we conducted mixed-design ANOVAs with serial position as a within-participant factor and cue type as a between-participant factor on the parameter estimates. None of the ANOVA effects were statistically significant for any modeled outputs (Table 2). However, the model outputs, particularly target rate, swap rate, and guess rate, significantly deviated from a normal distribution (Shapiro-Wilk test, ps < 0.001 for quadrant cue). Therefore, we conducted an additional mixed-design ART-ANOVA on these parameters.

Table 2.

ANOVA on modeled outputs for Experiment 2.

	Cue	Serial position	Cue × Serial position
Memory imprecision
ANOVA	F(1,25) = 0.57, p = 0.46, ${\eta }_{\mathrm{p}}^2$ = 0.02	F(3,75) = 1.02, p = 0.39, ${\eta }_{\mathrm{p}}^2$ = 0.04	F(3,75) = 0.28, p = 0.84, ${\eta }_{\mathrm{p}}^2$ = 0.01
ART-ANOVA	F(1,25) = 0.48, p = 0.50, ${\eta }_{\mathrm{p}}^2$ = 0.02	F(3,75) = 1.92, p = 0.13, ${\eta }_{\mathrm{p}}^2$ = 0.07	F(3,75) = 0.16, p = 0.92, ${\eta }_{\mathrm{p}}^2$ = 0.006
Target rate
ANOVA	F(1,25) = 3.02, p = 0.10, ${\eta }_{\mathrm{p}}^2$ = 0.11	F(3,75) = 1.76, p = 0.16, ${\eta }_{\mathrm{p}}^2$ = 0.07	F(3,75) = 2.61, p = 0.06, ${\eta }_{\mathrm{p}}^2$ = 0.10
ART-ANOVA	F(1,25) = 18.44, p < 0.001, ${\eta }_{\mathrm{p}}^2$ = 0.42***	F(3,75) = 2.82, p = 0.04, ${\eta }_{\mathrm{p}}^2$ = 0.10*	F(3,75) = 4.61, p = 0.005, ${\eta }_{\mathrm{p}}^2$ = 0.16**
Swap rate
ANOVA	F(1,25) = 3.81, p = 0.06, ${\eta }_{\mathrm{p}}^2$ = 0.13	F(2.1, 53.5) = 1.94, p = 0.13, ${\eta }_{\mathrm{p}}^2$ = 0.07	F(2.1, 53.5) = 2.20, p = 0.12, ${\eta }_{\mathrm{p}}^2$ = 0.08
ART-ANOVA	F(1,25) = 28.44, p < 0.001, ${\eta }_{\mathrm{p}}^2$ = 0.53***	F(3,75) = 6.89, p < 0.001, ${\eta }_{\mathrm{p}}^2$ = 0.22***	F(3,75) = 8.41, p < 0.001, ${\eta }_{\mathrm{p}}^2$ = 0.25***
Guess rate
ANOVA	F(1,25) = 0.98, p = 0.33, ${\eta }_{\mathrm{p}}^2$ = 0.04	F(1.7,43.1) = 0.62, p = 0.61, ${\eta }_{\mathrm{p}}^2$ = 0.02	F(1.7, 43.1) = 0.33, p = 0.81, ${\eta }_{\mathrm{p}}^2$ = 0.01
ART-ANOVA	F(1,25) = 0.49, p = 0.49, ${\eta }_{\mathrm{p}}^2$ = 0.02	F(3,75) = 0.48, p = 0.69, ${\eta }_{\mathrm{p}}^2$ = 0.02	F(3,75) = 0.76, p = 0.51, ${\eta }_{\mathrm{p}}^2$ = 0.03

*p < 0.05. **p < 0.01. ***p < 0.001.

The mixed-design ART-ANOVA revealed significant effects of cue type, serial position, and their interaction on the swap rate (Fig. 5B) and the target rate (Fig. 5C), but not on memory imprecision (Fig. 5A) or the guess rate (Fig. 5D; Table 2). Specifically, for the swap rate, there were significant effects of cue type, F(1,25) = 28.44, p < 0.001, serial position, F(3,75) = 6.89, p < 0.001, and their interaction, F(3,75) = 8.41, p < 0.001.

Figure 5.

(A–F) The swap model results and transposition error of memory-guided saccades in Experiment 2. Figure notations are the same as Fig. 3.

Similar to Experiment 1, the transposition error of the primary saccade (Table 2; Figs. 5E, 5F) varied significantly by cue type, F(1,25) = 22.76, p < 0.001, ${\eta }_{\mathrm{p}}^2$ = 0.48, and serial position, F(3,75) = 4.87, p = 0.004, ${\eta }_{\mathrm{p}}^2$ = 0.16. The interaction between serial position and cue type approached significance, F(3, 75) = 2.49, p = 0.07, ${\eta }_{\mathrm{p}}^2$ = 0.09. Specifically, transposition error was significantly higher during the order cue than the quadrant-cue condition, t(25) = 4.77, p < 0.001, Cohen's d = 1.38. Significant serial-position effects on transposition error were observed for the order cue, F(3, 30) = 3.20, p = 0.04, ${\eta }_{\mathrm{p}}^2$ = 0.24, with the linear contrast reaching significance, t(30) = −2.84, p = 0.008. No serial-position effects were observed in the quadrant-cue condition, F(3, 45) = 1.04, p = 0.38, ${\eta }_{\mathrm{p}}^2$ = 0.07.

Summary of results

Collectively, Experiment 2 largely replicated the findings of Experiment 1 despite differences in design and sample. Across both experiments, serial-position effects were significant in the order-cue but not the quadrant-cue condition, for both the non-modeled measures (endpoint error, latency, and transposition error) and the modeled swap rate. However, two differences emerged. First, the primacy effect in saccade error did not reach significance in Experiment 2, though the effect size was comparable with that in Experiment 1 (position 1 vs. 2: Cohen's d = −0.30 in experiment 1 vs. −0.37 in experiment 2). Second, cue-type effects on saccade error were less robust in Experiment 2, reaching significance only for transposition error and modeled swap rate. This weaker cue-related effect in Experiment 2 may reflect the between-participant manipulation of cue type, in contrast with the within-participant manipulation in Experiment 1.

Discussion

Across the two experiments, we found that serial-position effects in visuospatial working memory were more pronounced when participants maintained and bound two features (spatial location and serial order) compared with when only spatial location was task relevant. In the quadrant-cue condition, where serial order was present during encoding but not explicitly required for recall, saccade error, latency, and modeled memory precision were largely uniform across serial positions. In the order-cue condition, where successful retrieval depended on binding spatial location to serial position, robust primacy and recency effects emerged (Figs. 2 and 4). Statistical modeling revealed that the order-cue data were best fit by a swap model, whereas the quadrant-cue data were best fit by a non-swap model, with no clear serial-position modulation of the swap rate or transposition error (Figs. 3 and 5). These results suggest that serial-position effects in visuospatial working memory are primarily driven by misbinding between spatial location and serial position, rather than by changes in spatial working memory precision.

Feature misbinding as a major source of serial-position effects

A central implication of our findings is that serial-position effects primarily reflect failures of feature binding. This interpretation is consistent with studies of delayed recall of orientation (Gorgoraptis et al., 2011) and motion (Zokaei et al., 2011), where sequential presentation selectively increased modeled swap errors without changing precision compared with simultaneous presentation. However, because those paradigms required maintaining multiple visual features (e.g., orientation + color), it was difficult to isolate binding demands from feature-level precision. We addressed this issue by directly contrasting a multi-feature condition (order cue: location + ordinal position) with a single-feature condition (quadrant cue: location only). Across the model-free and model-based analyses in our study, serial-position effects were consistently stronger in the multi-feature condition and were accounted for by more swap rates and transposition errors rather than memory precision, providing direct evidence that feature misbinding is a key driver of serial-position effects.

Our findings diverge from McAteer et al. (2023), who reported no serial-position modulation of swap errors in a similar paradigm, a discrepancy likely attributable to their limited sample size (n = 10 in the partial-report condition).

More broadly, our results align with the temporal binding model of serial-order working memory (Schneegans et al., 2023), which posits that visual features are bound to ordinal positions rather than directly to each other. Schneegans et al. (2023) found stronger spatial-ordinal binding than spatial-color binding and observed correlated swap errors between visual features (shape or color) and ordinal position. Our data are consistent with this framework, confirming that instability in spatial-ordinal binding is a key mechanism underlying serial-position effects in visuospatial working memory.

Serial-position effects on memory precision

Contrary to our initial hypothesis, we did not observe robust serial-position effects on spatial memory precision. In the quadrant-cue condition, where serial order was present during encoding but not explicitly required for recall, neither saccade error nor latency varied significantly across serial positions. Mixture modeling revealed no reliable serial-position modulation of memory precision in either cue condition. These findings align with partial-report studies showing no significant serial-position effects on spatial or motion working memory precision (McAteer et al., 2023, experiment 3; Zokaei et al., 2011, experiment 1), although they contrast with Gorgoraptis et al. (2011), who reported serial-position effects on both swap rates and precision. Notably, in that study, only swap rates, and not precision, distinguished sequential from simultaneous presentation, reinforcing the central role of misbinding in serial-order processing.

Importantly, the absence of strong serial-position effects on precision does not imply that such effects cannot occur. Proactive and retroactive interference between memory traces could, in principle, produce graded changes in representational fidelity across serial positions. However, as our results suggest, serial-position effects on memory precision are likely smaller in magnitude than feature-binding effects and may emerge primarily under specific task conditions. For example, primacy or recency effects for spatial location recall have been observed with longer lists (six or more items) in free-recall paradigms (Bonanni et al., 2007; Cortis et al., 2015), potentially reflecting increased interference as capacity limits are approached or additional output interference during retrieval. Hay, Smyth, Hitch, and Horton (2007) further showed that stimulus distinctiveness modulates serial-position profiles in single-feature precision tasks: highly distinctive stimuli (e.g., upright faces) produced extended recency effects, whereas less distinctive stimuli (e.g., random matrices) yielded flatter curves. Task structure may also play a critical role. In our design, set size was fixed and probe probability was equal across items, encouraging a more even allocation of mnemonic resources across serial positions. In contrast, tasks with variable or unpredictable set sizes may promote prioritization strategies that increase serial-position differences in item recall (Gorgoraptis et al., 2011). Together, these findings suggest that serial-position effects on precision are relatively unstable and highly sensitive to task parameters.

By contrast, precision appears to be more consistently influenced by set size than by serial order. A large body of work shows that memory precision declines reliably as load increases (Bays & Husain, 2008; Gorgoraptis et al., 2011; Zokaei et al., 2011), whereas serial-position effects on precision are less stable and more task dependent. Our findings indicate that, under the present spatial working memory design, serial-position effects are more plausibly driven by feature-binding processes than by systematic changes in spatial memory precision. Future work that systematically manipulates list length, stimulus similarity, output demands, and task predictability will be necessary to delineate the conditions under which serial-position effects on precision reliably emerge.

Robust recency and variable primacy effects

We observed consistent recency effects across both experiments, but more variable primacy effects. The recency effect, indexed by the difference between serial positions 2 and 4, was significant in both experiments. The primacy effect (position 1 vs. position2) reached statistical significance threshold only in Experiment 1. In Experiment 2, the trend was in the expected direction, with a comparable effect size, but did not attain statistical significance. This pattern is consistent with partial-report studies reporting robust recency alongside weaker primacy (Gorgoraptis et al., 2011; Hay et al., 2007; McAteer et al., 2023; Zokaei et al., 2011), where the attenuated primacy likely reflects the absence of output interference and reduced strategic rehearsal relative to full-report designs (McAteer et al., 2023). Additionally, our longer retention intervals (4–10 seconds in Experiment 1; 4–6 seconds in Experiment 2) compared with previous partial-report studies (0.5–1.0 second) (Gorgoraptis et al., 2011; McAteer et al., 2023; Zokaei et al., 2011) may have contributed to the primacy effects observed in Experiment 1, because longer delays may increase the temporal distinctiveness of early items (Brown, Neath, & Chater, 2007; Knoedler, Hellwig, & Neath, 1999; Neath, 1993).

Serial-position effects on different saccade metrics

Our eye-tracking approach offered insights into how serial-position effects vary across different types of eye movements. We observed a strong interaction between serial position and cue condition for primary saccades, whereas closest saccades to the target showed smaller and comparable serial-position modulation across cue conditions (Fig. 2), paralleling the contrast between modeled swap rate and memory precision (Fig. 3). Possibly, primary memory-guided saccades reflect an initial, crude memory retrieval that is more susceptible to item swapping, whereas closest saccades may reflect a more refined memory representation related to memory precision for the remembered location.

Analyses of corrective saccades provide complementary insight into serial-position effects on saccadic behavior. Intriguingly, corrective saccade rates varied with serial position, but only in the order-cue condition (Supplementary Fig. S5). This pattern partially resembles the serial-position effects observed for the primary saccade error and the modeled swap rate, both of which were modulated only in the order-cue condition, but differs in shape: corrective saccade rates increased at later positions, whereas the primary saccade error and swap rate were highest for intermediate positions. When one-step (no corrective saccades) and multi-step (requiring corrective saccades) trials were analyzed separately, the closest-saccade error showed no significant serial-position effects or Serial-position × Cue interactions in either trial type (Supplementary Fig. S6). This finding indicates that serial position modulated the frequency of making corrective eye movements, but could not completely correct the memory-guided saccade error, although lower trial counts in these split analyses may have limited our ability to detect such effects.

Corrective saccades did, however, partially compensate for cue-related differences in rates of item swapping. Further decomposing the transposition errors trials revealed that both correctable transpositions (primary saccade in a non-target quadrant but with a subsequent saccade reaching the target) and uncorrectable transpositions (the subsequent saccade also failed reaching the target) occurred more often during the order-cue condition relative to the quadrant-cue condition. However, the cue-related increase in the swap rate was disproportionately more frequent for uncorrectable errors (Supplementary Fig. S7), suggesting that, although some swap errors may reflect transient retrieval difficulties that corrective saccades can resolve, item swapping more often links to genuine memory misbinding, which cannot be fully corrected.

Finally, whereas previous studies have primarily focused on examining response accuracy across serial positions, we found robust primacy and recency effects on saccade latency in the order-cue condition that were absent in the quadrant-cue condition. These results suggest that saccade latency might be a robust indicator of serial-order processing. Specifically, the serial-position curve of latency closely resembles that of endpoint error, transposition error, and modeled swap rate of primary saccades, indicating that latency is a sensitive measure that may closely reflect the misbinding between spatial location and serial position.

Limitations and alternative hypotheses

Several limitations and alternative hypotheses warrant consideration. First, the order-cue condition imposed both greater binding requirements and potentially a greater memory load, raising the possibility that stronger serial-position effects may simply reflect increased difficulty. However, saccade error did not consistently differ between cue conditions across experiments—only in Experiment 1 and only for primary saccades—whereas the Cue × Serial-position interaction was robust across measures and experiments. If difficulty alone drove the effect, a consistent main effect of cue would be expected. Instead, the reliable interaction between cue and serial-position across experiments argues against a pure difficulty account and instead supports a role for binding demands in the reported findings.

Second, the spatial nature of the quadrant cue may have facilitated more accurate yet less differentiated responses across serial positions. However, it is worth mentioning that, in our study, we cued location with spatial context (quadrant membership), not spatial locations. That means that participants would still need to remember and report the exact location of the target, and knowing the quadrant does not guarantee precise memory recall. For example, our empirical results showed that the precision of the closest saccade toward the target was comparable between the quadrant-cue and order-cue conditions across both experiments. Moreover, when cue type was manipulated between subjects in Experiment 2, even primary saccade errors were similar across the two cue conditions. These findings are inconsistent with the interpretation that quadrant cueing would artificially inflate memory precision performance. Furthermore, prior work suggests that categorical (e.g., quadrant membership) and coordinate (precise metric location) spatial representations rely on at least partially dissociable cognitive and neural systems (Van Der Ham, Raemaekers, Van Wezel, Oleksiak, & Postma, 2009; Van Der Ham et al., 2012). Categorical spatial judgments are more strongly associated with left parietal and premotor regions and may engage verbal or symbolic coding, whereas coordinate spatial memory depends more on right parietal cortex and retinotopic representations, supporting metric precision. Thus, although both cue and report features are spatial, they likely engage at least partially dissociable representational formats.

Third, mixture-model analysis raises issues of parameter identifiability. Parameters such as memory imprecision and guess are not fully independent and may trade off during fitting. To mitigate this, we constrained the uniform (guessing) distribution to the stimulus space while allowing Gaussian components to remain unbounded, improving separability. Model comparisons using both the BIC and the AIC favored the swap model in the order-cue condition but simpler models for the quadrant-cue condition, indicating that the additional swap parameter contributed meaningful explanatory power.

Finally, our design used only horizontal sequences (left-to-right or right-to-left), which may encourage strategic chunking or relational encoding (Rosenbaum, Kenny, & Derr, 1983) and influence sequence reproduction accuracy (Zhang et al., 2022). Excluding vertical sequences is our deliberate design choice to reduce design complexity and to study serial-position effects under a well-controlled spatial configuration. Notably, vertical and horizontal visual fields are subject to different perceptual asymmetries (Barbot, Xue, & Carrasco, 2021) and gravity-related biases (Deeb & Domini, 2024). Incorporating vertical sequences would potentially introduce additional variability and confounds that could further complicate date interpretation. Nonetheless, examining whether serial-position effects generalize across spatial configurations and relational structures represents an important direction for future research.

Conclusions

Taken together, our study suggests that feature misbinding is a primary driver of serial-position effects in visuospatial working memory. When only spatial location was task relevant and serial position was not explicitly required at recall, items were retrieved with comparable accuracy and latency across serial positions. However, when successful retrieval depended on binding spatial location to serial position, serial-position effects emerged, with intermediate items recalled less precisely and more slowly than earlier and later items in the sequence. These findings indicate that limitations in serial-order working memory are shaped by the capacity to maintain bound, associative representations over time.