Human cognition is capacity-limited, requiring strategies to actively structure information. Eye movements offer a natural mechanism for sequential sampling, but whether such sequences reflect memory organization is unknown. We developed a working-memory task where color-frequency pairings created a latent ordinal structure to optimally reduce memory load. Across two experiments, gaze patterns spontaneously aligned with this structure. Participants sampled items following this sequence during encoding and revisited them during maintenance. Critically, the expression of this structure depended on cognitive demand. In a 3-item task, high performers showed robust sequential sampling during encoding, whereas lower performers compensated with replay-like revisitation during maintenance. Under higher demand (4 items), encoding-based organization was disrupted, and structured revisitation emerged primarily during maintenance. These findings suggest that eye movements provide an overt expression of an internally generated organizational process that flexibly adapts to cognitive demands, revealing a behavioral analogue to neural replay when encoding resources are strained.

Eye-movement trajectory is modulated by the spatial configuration of the memory contents

Participants were briefly shown three gratings presented simultaneously at distinct spatial locations, each varying in spatial frequency and color. At the end of each trial, they reported the location of a probed feature (either spatial frequency or color) or indicated that the feature was new (Fig 1A). This task challenged working memory capacity by requiring the maintenance of multiple feature–location bindings. To probe potential efficient coding strategies, we introduced an implicit regularity in the stimulus set: lower-frequency gratings were consistently paired with “greener” hues. This contingency established an ordinal structure among the items, providing a potential scaffold for more efficient memory organization. Crucially, on every trial, items could be ranked according to their relative position along this feature gradient (item 1: lowest frequency/greenest; item 2: intermediate; item 3: highest frequency/least green). These ordinally labeled items were assigned to three fixed locations, yielding six possible configuration patterns (Fig 1B). To encourage natural encoding strategies, participants were allowed to move their eyes freely except when a central fixation dot was presented, during which they were required to maintain fixation. Our analyses focused on eye-movement patterns during both the encoding period (when the memory items were visible) and the maintenance period (after their disappearance).

Behavioral performance was well above chance level (0.25) across all conditions (overall: mean = 0.66, SD = 0.12, t₍₄₅₎ = 22.46, p < 0.001, Cohen’s d = 3.31; color domain: mean = 0.70, SD = 0.13, t₍₄₅₎ = 22.87, p < 0.001, Cohen’s d = 3.37; frequency domain: mean = 0.62, SD = 0.16, t₍₄₅₎ = 15.50, p < 0.001, Cohen’s d = 2.29; color performance is significantly higher: t₍₄₅₎ = 3.41, p = 0.001, Cohen’s d = 0.50). We then asked whether eye-movement trajectories systematically reflected the embedded cross-feature ordinal structure. Specifically, if participants sampled information based on this latent item order, eye-movement trajectories would be more similar for trials sharing the same configuration (within-pattern) than for trials with different configurations (between-pattern). To test this configuration-specific pattern-similarity effect, we quantified attentional allocation using normalized gaze-to-item distance time courses (Fig 2A), with larger values reflecting stronger relative gaze bias toward a given item. For each item, we computed pairwise correlations between these normalized distance time courses across trials and categorized each pair as within-pattern or between-pattern depending on whether the two trials shared the same configuration or not. Correlations were then averaged across items to yield a single summary measure per participant (see Methods). The difference between within- and between-pattern correlations quantified the pattern-similarity effect, providing an index of structured gaze organization while controlling for baseline biases associated with spatial location.

As predicted, we observed a significant pattern-similarity effect, i.e., gaze trajectories were more similar within-pattern than between-pattern during both the encoding (t₍₄₅₎ = 7.77, p < 0.001, Cohen’s d = 1.15) and maintenance period (t₍₄₅₎ = 4.15, p < 0.001, Cohen’s d = 0.61, Fig 1C). This effect remained when we balanced trial numbers across within- and between-pairs (see S1A Fig) and was already presented in the first half of the trials (see S4 Fig). Critically, participants who showed a stronger pattern-similarity effect during encoding also exhibited higher memory accuracy (r = 0.38, p = 0.009, Fig 1D). This suggests that stronger configuration-specific organization of gaze during encoding reflects more structured mnemonic processing associated with better memory performance.

Sequential eye movements reveal performance-dependent sampling and revisitation

Having established that gaze patterns reflect structured mnemonic processing, we next examined the temporal organization of these trajectories. Specifically, we asked whether participants sampled items in a systematic sequence that strictly followed their ordinal structure (either 1 → 2 → 3 or 3 → 2 → 1, where 1–3 correspond to items ordered from lowest to highest spatial frequency). To quantify sequential structure, we applied a sequenceness analysis framework [21]. We first computed cross-correlation coefficients between all pairs of item-specific gaze-bias time courses to generate a correlation matrix for each temporal lag (Fig 2B). Next, we specifically tested two transition models against these empirical matrices: rank-increasing sequence 1 → 2 → 3 model which contained transitions from item 1 to item 2 and from item 2 to item 3, and a reverse 3 → 2 → 1 model (Fig 2C, see Methods). Permutation tests with item-label shuffling showed that only the 1 → 2 → 3 sequence significantly explained the empirical data, with a prominent effect at lags of 0.75–1.07 s (peak = 0.91 s, Fig 2D, left). To assess behavioral relevance, we split participants into high- and low-performance groups based on a median split of memory accuracy (N = 23 per group). Strikingly, the high-performance group exhibited a robust 1 → 2 → 3 sequential profile (lags: 0.63–1.17 s; peak = 0.89 s), whereas the low-performance group did not (Fig 2D, right), indicating that ordered gaze sampling during encoding was more strongly expressed in participants with better memory performance. This sequential pattern also became more strongly expressed with experience across trials (see S4 Fig) and remained visible when looking at discrete fixation sequence (see S7I Fig).

Having characterized the sequential structure during encoding, we turned to the maintenance period to determine whether the same sequence re-emerged after the stimuli were no longer visible. To address this, we applied the same sequenceness analysis to the maintenance period. At the group level, no significant sequential eye-movement pattern was observed. However, analysis of individual differences revealed a striking and unexpected dissociation: contrary to what might be assumed, the low-performance group exhibited significant sequential “revisitation” at lags of 0.68–0.95 s (peak: 0.72 s), whereas the high-performance group showed no evidence. Importantly, control analyses confirmed this effect was not explained by spatial biases, as sequence analysis based solely on item locations yielded no significant structure (S2 Fig).

Taken together, these findings show that structured eye-movement patterns were closely associated with memory performance and differed between encoding and maintenance as a function of individual ability. High-performing participants showed robust ordinal sampling (lowest → highest) during encoding but no evidence of sequential revisitation during maintenance, consistent with the idea that efficient initial encoding reduces the need for subsequent stabilization. In contrast, low-performing participants failed to deploy this encoding strategy and instead exhibited revisitation during maintenance, consistent with post-encoding reorganization and stabilization of memory traces. This behavioral pattern aligns with evidence from neural replay, where replay primarily supports the consolidation of weakly encoded information, whereas strong initial encoding diminishes the need for later reactivation [22,23].

Retro-cue initiates sequential gaze transitions following the ordinal structure

To further validate the observed sequential organization, we examined eye-movement sequence following a retro-cue presented immediately after the maintenance period. The retro-cue was designed to exogenously direct attention and gaze. If effective, the initial saccade should be directed toward the cued item, whereas subsequent saccades may reveal the internal priority structure of the remaining memory representations. Given the observed 1 → 2 → 3 ordering, we predicted that: (1) when item 1 was cued, the second saccade would be more likely to target item 2 than item 3, and (2) when item 2 was cued, the second saccade would be more likely to target item 3 than item 1. We first verified the effectiveness of the retro-cue by analyzing first-saccade destinations (Fig 3A). Fixation heatmaps following cue onset revealed a strong spatial bias toward the cued locations (Fig 3B). Quantitatively, first saccades were directed toward the cued location significantly more often than expected by chance (1/3; t₍₄₅₎ = 7.19, p < 0.001, Cohen’s d = 1.06; Fig 3C).

To test whether the second saccade (Fig 3D) followed the predicted sequential order, we focused on trials that met three criteria: (1) the first saccade landed on the cued location, confirming effective cueing; (2) either item 1 or item 2 (the positions from which ordered transitions are interpretable) was cued; and (3) at least two distinct saccades were made. Note that 4 participants were excluded from further analyses because fewer than 20 trials satisfied these requirements. Consistent with our hypothesis, second-saccade transitions exhibited strong directionality: item 1 → item 2 and item 2 → item 3 transitions occurred significantly more frequently than alternative transitions (item 1 → item 3 and item 2 → item 1) (t₍₄₁₎ = 3.33, p = 0.002, Cohen’s d = 0.51; Fig 3E), which robustly validated the optimal 1 → 2 → 3 sequential organization. Additional analyses including all six possible transitions (1 → 2, 1 → 3, 2 → 1, 2 → 3, 3 → 1, 3 → 2) yielded a similar pattern. Forward transitions (1 → 2, 2 → 3) occurred significantly more often than backward transitions (2 → 1, 3 → 2) (t₍₄₁₎ = 2.36, p = 0.04, Cohen’s d = 0.32) and long-distance transition (1 → 3, 3 → 1) (t₍₄₁₎ = 3.35, p = 0.004, Cohen’s d = 0.46) (S5B Fig). Together, these results show that post-maintenance gaze transitions precisely mirror the same ordinal structure observed during encoding/maintenance, confirming the active sequential organization of memory representation.

Central-fixation constraint degraded memory performance

The association between structured gaze and memory performance in Experiment 1 raised an important question: what happens to memory organization when these natural eye-movement patterns are restricted? To address this question, we conducted a control experiment using the same 3-item task as in Experiment 1, except that participants were required to maintain central fixation during encoding and maintenance (Fig 4A). Compared with the free-viewing condition, memory accuracy (mean = 0.55, SD = 0.09) was lower under this fixation constraint (Fig 4B, t₍₇₅₎ = 4.37, p < 0.001, Cohen’s d = 1.01). Moreover, the configuration-specific pattern-similarity effect observed under free viewing was no longer significant under the fixation constraint (Fig 4C, encoding period: t₍₃₀₎ = 1.44, p = 0.16, Cohen’s d = 0.26; maintenance period: t₍₃₀₎ = –0.43, p = 0.67, Cohen’s d = –0.08), nor was the corresponding correlation with memory performance (Fig 4D, r = 0.27, p = 0.15). Together, these findings are consistent with the view that overt eye movements reflect organization of memory representations.

High cognitive load shifts structured organization from encoding to maintenance

Our previous results suggested that structured gaze sequences are closely linked to successful memory organization, but that their expression may depend on available cognitive resources. High performers showed robust sequential sampling during encoding, whereas low performers appeared to shift this organizational process into the maintenance period, where sequential revisitation may reflect post-encoding reorganization of less robustly encoded information. This pattern raised a key question: Does increasing cognitive load disrupt the expression of sequential organization during encoding and promote a compensatory shift toward structured revisitation during maintenance? To address this, we designed a higher-load 4-item memory task (Fig 5A), analogous to the 3-item paradigm but with one additional item to remember.

Participants performed well above chance (0.2; overall: mean = 0.59, SD = 0.11, t₍₃₉₎ = 22.33, p < 0.001, Cohen’s d = 3.53 (Fig 5B right); color domain: mean = 0.61, SD = 0.12, t₃₉₎ = 21.05, p < 0.001, Cohen’s d = 3.33; frequency domain: mean = 0.57, SD = 0.14, t₍₃₉₎ = 17.15, p < 0.001, Cohen’s d = 2.71; performance in color domain is marginally higher than in the frequency domain: t(39) = 1.73, p = 0.09, Cohen’s d = 0.27). Directly comparing the memory accuracy between the 3-item and 4-item tasks was not appropriate because they differed in chance level (0.25 versus 0.2). We therefore compared performance using chance-corrected accuracy, , where is observed accuracy and is the chance level. Under this normalization, performance in the 4-item task (mean = 0.49, SD = 0.14) was lower than in the 3-item task (mean = 0.54, SD = 0.16) (t₍₈₄₎ = 1.62, p = 0.05, one-tailed, Cohen’s d = 0.35), although the statistical evidence for this difference was modest. Nevertheless, the 4-item task imposed greater cognitive demands by design, as participants were required to maintain one additional item in working memory. The chance-corrected behavioral comparison provides convergent, albeit modest, support for this increased difficulty.

Replicating our 3-item analysis, we first examined whether eye-movement trajectories also showed a pattern-similarity effect, i.e., greater similarity for trials sharing the same configuration pattern relative to trials with different patterns. This pattern-similarity effect was again significant during both encoding (t₍₃₉₎ = 10.66, p < 0.001, Cohen’s d = 1.69) and maintenance (t₍₃₉₎ = 4.54, p < 0.001, Cohen’s d = 0.72, Fig 5C), with the encoding effect strongly associated with memory accuracy (r = 0.57, p < 0.001, Fig 5D). Critically, comparison of the 3-item and 4-item tasks revealed a significant interaction effect (Fig 5E, F_(1,84) = 6.84, p = 0.01, η_p² = 0.08): compared with the 3-item task, the 4-item task showed a marginally reduced pattern-similarity effect during encoding (t₍₈₄₎ = −1.86, p = 0.07, Cohen’s d = 0.40) but a significantly stronger effect during maintenance (t₍₈₄₎ = 2.35, p = 0.02, Cohen’s d = 0.51). These findings suggest that increasing task demand was associated with a shift in configuration-specific gaze organization across processing stages, with relatively weaker expression during encoding but stronger expression during maintenance.

We next asked whether the more fine-grained sequential structure of eye movements showed a comparable shift. Following the analysis in the 3-item task, we constructed 4 × 4 time-lagged cross-correlation matrices and test whether they reflected the 1 → 2 → 3 → 4 sequence (Fig 5F). During encoding, no significant group-level expression of this sequential pattern was observed (Fig 5G). Although this may appear inconsistent with the encoding-period pattern-similarity effect, the two analyses address different levels of structure. The pattern-similarity analysis asks whether participants exhibit similar configuration-specific gaze patterns across trials, whereas the sequenceness analysis additionally requires these gaze patterns to follow a stable ordinal trajectory over time.

We then again divided participants into high- and low-performance groups based on a median split of memory accuracy (N = 20 per group). Although high-performing participants showed a trend toward structured sampling during encoding, this effect did not reach statistical significance (see S3 Fig for nonsignificant 4 → 3 → 2 → 1 sequence). In contrast, during maintenance (Fig 5H), although no significant sequence was observed at the group level, high performers exhibited significant sequential revisitation at lags of 0.35–0.74 s (peak at 0.55 s). This result resembled the maintenance-period revisitation pattern observed in lower performers in the 3-item task, consistent with the idea that increasing cognitive demand shifts the expression of sequential organization toward maintenance, where internally generated revisitation may reflect compensatory reorganization of structured memory representations.

Taken together, our experiments demonstrate that eye movements reflect adaptive memory organization according to the latent ordinal structure. Critically, the deployment of this organizational strategy appears to depend on available cognitive resources. In the 3-item task, where encoding demands were relatively lower, high performers engaged in structured sampling during encoding, whereas low performers exhibited sequential revisitation during maintenance, consistent with compensatory post-encoding reorganization. In the more demanding 4-item task, encoding-phase sequential organization was less reliably expressed, while high performers developed structured revisitation during maintenance. Thus, under higher cognitive load, sequential organization appears to shift toward the maintenance period, potentially providing an opportunity to reorganize less robustly encoded information. More broadly, these findings reveal sequential eye movements as an overt expression of an active, adaptive mechanism for internally organizing information in working memory.

Participants

One hundred twenty-seven participants (aged 18–35 years) took part in the study: 48 in Experiment 1 (3-item memory), 38 in the central-fixation control experiment, and 41 in Experiment 2 (4-item memory). They were recruited via the internal database recruitment system of Max Planck Institute for Human Cognitive and Brain Sciences in Leipzig. The study was approved by the local Ethics Committee of Leipzig University, Germany (Ethics Approval: Nr 018/23-ek), and conducted in accordance with the principles expressed in the Declaration of Helsinki. All the participants were naïve to the purpose of the experiments and gave written informed consent before the experiment and were compensated for their participation with 13 Euros per hour. All participants had normal or corrected-to-normal vision, with no history of psychiatric or neurological disorders. Due to poor eye-tracking data quality, two participants were excluded from Experiment 1 and one from Experiment 2. In the central-fixation control experiment, one participant failed to complete the task and was excluded. For the remaining 37 participants, fixation compliance was evaluated so that only 31 participants who followed the central-fixation instruction sufficiently well were retained for further analysis.

Stimuli and tasks

Visual stimuli consisted of either three (Experiment 1 and fixation control experiment) or four (Experiment 2) items presented simultaneously at different locations (6° visual angle eccentricity). Each item (2.8° visual angle in radius) comprised two integrated components: (1) an internal grating with one of six possible spatial frequencies (0.6, 0.9, 1.3, 1.9, 2.7, and 4 cycles per visual degree, logarithmically spaced), and (2) an outer ring in one of six colors (evenly spaced from blue to green in Commission Internationale de l’Eclairage (CIE) LAB color space). A systematic relationship was established between stimulus features: lower spatial frequencies were consistently paired with greener hues, whereas higher spatial frequencies were paired with bluer hues. This created an implicit feature-space structure where color and frequency varied coherently along a single dimension. We used two feature dimensions that varied in a consistent ordinal manner because this creates a richer latent relational structure than either feature alone. Specifically, the shared ordinal relationship is expressed simultaneously across both color and spatial frequency, encouraging participants to represent the stimulus set in terms of its underlying structure rather than as two independent feature dimensions. In this sense, the latent structure is not tied to a particular feature, but to the ordinal relationship shared across features. We reasoned that such a representation would make relational encoding and ordinal sampling more advantageous than memorizing each feature independently, thereby increasing the likelihood that spontaneous structure-based sampling strategies would be expressed in gaze behavior.

Eye-tracking recordings

Participants sat 60 cm from a monitor (1,280 × 1,024 pixels, 60 Hz refresh rate) with their head stabilized using a chin rest. An eye tracker (EyeLink 1000 Plus, SR Research) was positioned on the table approximately 15 cm in front of the monitor. During the task, gaze was continuously tracked for both eyes simultaneously, at a sampling rate of 1,000 Hz. Before acquisition, we calibrated the eye tracker using the built-in calibration and validation protocols from the EyeLink software.

Eye-tracking preprocessing

Raw eye-tracking data were converted from EDF to MAT format using the Edf2Mat toolbox with default settings. For each trial, we extracted three epochs of interest: (1) the stimulus presentation epoch (−10 to 4,000 ms from stimulus onset), (2) the memory maintenance epoch (−10 to 2,000 ms from placeholder onset), and (3) the memory retrieval epoch (−10 to 2,100 ms from retro-cue onset). Baseline correction was applied using the pre-event interval (−10 to 0 ms). Note that the implementation of baseline correction was to reduce systematic trial-to-trial offsets in gaze position, because the gaze-distance analysis depends on accurately quantifying the spatial relationship between gaze position and the item locations. For each critical period of interest (encoding, maintenance, and retro-cue), we used the immediately preceding fixation interval as the baseline. During these intervals, participants were instructed to fixate at the center of the screen within 1 degree of visual angle from the center for at least 10 ms. Blinks were detected by identifying NaN clusters (≥9 consecutive missing samples) in each epoch. To capture pre/post-blink distortions, we extended each blink window by 100 ms on both sides. Gaze data from both eyes were then averaged, producing horizontal (x) and vertical (y) gaze time series per epoch. For data quality control, epochs with more than 50% invalid samples were discarded. Remaining blink artifacts were interpolated using MATLAB’s makima method (a smooth cubic Hermite interpolation). Finally, trials where gaze deviated >10° visual angle from central fixation were excluded to ensure proper gaze. Following preprocessing, the mean number of valid trials retained in Experiment 1 was 198 during encoding (SD = 23, range = 108–216) and 196 during maintenance (SD = 25, range = 102–216). In Experiment 2, the mean number of valid trials was 186 during encoding (SD = 26, range = 108–214) and 184 during maintenance (SD = 25, range = 72–216).

Eye-tracking analysis

S1 Fig. (A) To account for unequal trial numbers between the within-pattern and between-pattern conditions in Experiment 1, we randomly subsampled the between-pattern trial pairs to match the number of within-pattern pairs.

This procedure was repeated 100 times, and the resulting correlations were averaged across iterations. The pattern similarity effect (mean ± SEM) was then computed separately for the encoding (yellow) and maintenance (purple) periods. *: p < 0.05. (B) To assess the behavioral relevance of the 3 → 2 → 1 sequence, we compared sequence strength (mean ± SEM) between high (dark red) and low-performance (light red) groups, defined by a median split of memory accuracy (N = 23 per group). Results are shown separately for encoding (left panel) and maintenance (right panel). The corresponding data are publicly available on OSF: https://osf.io/kasr2/.

https://doi.org/10.1371/journal.pbio.3004047.s001

(TIF)

S2 Fig. Six possible sequences based solely on spatial locations.

To rule out the possibility that the observed sequential eye-movement patterns were driven by systematic transitions between spatial locations rather than the latent ordinal structure, we relabeled items according to their spatial locations and tested all six possible spatial sequences during encoding (A–C) and maintenance (D–F), using the same sequenceness analysis. No significant sequential patterns were observed, indicating that the original sequential effects were not explained by spatial location alone. The corresponding data are publicly available on OSF: https://osf.io/kasr2/.

https://doi.org/10.1371/journal.pbio.3004047.s002

(TIF)

S3 Fig. Sequence strength of the reverse 4 → 3 → 2 → 1 ordinal sequence in Experiment 2.

The reverse ordinal sequence (4 → 3 → 2 → 1) was not significantly expressed during either encoding (A–Left) or maintenance (B–Left). To examine potential individual differences, participants were divided into high (dark red) and low-performance (light red) groups based on a median split of memory accuracy (N = 20 per group). No significant 4 → 3 → 2 → 1 sequence was observed in either group during encoding (A–Right) or maintenance (B–Right), providing no evidence for reliable expression of the reverse ordinal sequence at either the group or subgroup level. The corresponding data are publicly available on OSF: https://osf.io/kasr2/.

https://doi.org/10.1371/journal.pbio.3004047.s003

(TIF)

S4 Fig. Learning effects across early and late trial halves.

To test the learning effect, trials were split into early and late halves. Memory accuracy (A, G), pattern-similarity effect (B, H), sequential organization (C–F and I–L) were separately tested for the early and late halves for Experiment 1(3-item task) and Experiment 2 (4-item task). Error bars indicate SEM. Dots represent individual participants. *: p < 0.05. Memory performance (A, 3-item task; G, 4-item task). Accuracy was significantly higher in the late than in the early half in both tasks (3-item: t₍₄₅₎ = 4.57, p < 0.001, Cohen’s d = 0.67; 4-item: t₍₃₉₎ = 5.03, p < 0.001, Cohen’s d = 0.80). The pattern-similarity effect (B, 3-item task; H, 4-item task). In the 3-item task, the pattern-similarity effect was present during encoding in both the early and late halves (early: t₍₄₅₎ = 5.43, p < 0.001, Cohen’s d = 0.80; late: t₍₄₅₎ = 7.40, p < 0.001, Cohen’s d = 1.09). During maintenance, the effect was absent in the early half but emerged in the late half (early: t₍₄₅₎ = 1.34, p = 0.19, Cohen’s d = 0.20; late: t₍₄₅₎ = 2.90, p = 0.006, Cohen’s d = 0.43). In the 4-item task, the effect was observed in both halves during both encoding (early: t₍₃₉₎ = 8.43, p < 0.001, Cohen’s d = 1.33; late: t₍₃₉₎ = 8.50, p < 0.001, Cohen’s d = 1.34) and maintenance (early: t₍₃₉₎ = 3.03, p = 0.004, Cohen’s d = 0.48; late: t₍₃₉₎ = 2.97, p = 0.005, Cohen’s d = 0.47). A significant difference between the early and late halves was observed only during encoding in the 3-item task (t₍₄₅₎ = 5.23, p < 0.001, Cohen’s d = 0.77), but not during maintenance in the 3-item task (t₍₄₅₎ = 1.33, p = 0.19, Cohen’s d = 0.20), or during encoding or maintenance in the 4-item task (encoding: t₍₃₉₎ = 1.15, p = 0.26, Cohen’s d = 0.18; maintenance period: t₍₃₉₎ = 0.77, p = 0.45, Cohen’s d = 0.12). Sequential organization (C–F, 3-item task; I–L, 4-item task). At a descriptive level, the 1 → 2 → 3 sequence during encoding in the 3-item task was present as a trend in the early half, particularly in high-performing participants (C), and became more pronounced in the late half (D). Similarly, in the 4-item task, the 1 → 2 → 3 → 4 sequential revisitation during maintenance was more clearly expressed in the late half than in the early half, again particularly in high-performing participants (K, L). The corresponding data are publicly available on OSF: https://osf.io/kasr2/.

https://doi.org/10.1371/journal.pbio.3004047.s004

(TIF)

S5 Fig. (A) To test whether the forward bias might reflect anchoring on the low-frequency item (item 1), we examined the distribution of the first fixation during encoding.

Specifically, we quantified how often the first fixation fell closest to item 1 (yellow), item 2 (purple), or item 3 (green). The first fixations were directed significantly more often toward item 1 than expected by chance (chance = 72, t₍₄₅₎ = 2.06, p = 0.05, Cohen’s d = 0.30), whereas fixations toward item 2 (t₍₄₅₎ = –5.04, p < 0.001, Cohen’s d = –0.74) and item 3 (t₍₄₅₎ = –1.51, p = 0.14, Cohen’s d = –0.22) were less frequent. Direct comparisons further showed that first fixations toward item 1 were significantly more frequent than those toward item 2 (t₍₄₅₎ = 3.45, p = 0.001, Cohen’s d = 0.51) and marginally more frequent than those toward item 3 (t₍₄₅₎ = 1.86, p = 0.07, Cohen’s d = 0.28). (B) Following the retro-cue, six transitions were possible(1 → 2; 1 → 3; 2 → 1; 2 → 3; 3 → 1; 3 → 2). Transition frequencies were grouped into three types: forward transitions (1 → 2; 2 → 3, yellow), backward transitions (2 → 1; 3 → 2, purple), and long-distance transitions (1 → 3; 3 → 1, green). A significant main effect of transition type was observed (F_(2,82) = 5.91, p = 0.004, η_p² = 0.13). Post hoc comparisons showed that forward transitions occurred significantly more often than backward transitions (t₍₄₁₎ = 2.36, p = 0.04, Cohen’s d = 0.32) and long-distance transition (t₍₄₁₎ = 3.35, p = 0.004, Cohen’s d = 0.46), whereas backward and long-distance transitions did not differ significantly (t₍₄₁₎ = 0.99, p = 0.32, Cohen’s d = 0.14). Error bars indicate SEM. Dots represent individual participants. *: p < 0.05. The corresponding data are publicly available on OSF: https://osf.io/kasr2/.

https://doi.org/10.1371/journal.pbio.3004047.s005

(TIF)

S6 Fig. (A) First-saccade destination heatmaps for each cue location. Red dashed circles mark the cued locations.

(B) Number of trials containing two distinct saccades in 3-item (gray) and 4-item (black) tasks. As in Experiment 1, participants with fewer than 20 trials containing two distinct saccades were excluded from this analysis (four participants in Experiment 2). The number of trials containing two distinct saccades was significantly lower in Experiment 2 than in Experiment 1 (Fig B; t₍₇₆₎ = −3.54, p = 0.001, Cohen’s d = −0.80). (C) Number of trials containing forward (1 → 2; 2 → 3; 3 → 4) and backward transitions (2 → 1; 3 → 2; 4 → 3) in Experiment 2. Forward transitions did not occur significantly more frequently than backward transitions (Fig C, t₍₃₅₎ = −0.03, p = 0.98, Cohen’s d = −0.005), nor did they occur more often than expected by chance (t₍₃₅₎ = 0.79, p = 0.44, Cohen’s d = 0.13). Error bars indicate SEM. Dots represent individual participants. *: p < 0.05. The corresponding data are publicly available on OSF: https://osf.io/kasr2/.

https://doi.org/10.1371/journal.pbio.3004047.s006

(TIF)

S7 Fig. Basic eye-movement measures and fixation-sequence analysis in the 3-item and 4-item tasks.

(A) Number of saccades during encoding in the 3-item (yellow) and 4-item (purple) tasks. Participants made more saccades in the 4-item than in the 3-item task (t₍₈₄₎ = 6.94, p < 0.001, Cohen’s d = 1.50). (B) Number of fixations during encoding. More fixations were made in the 4-item compared to 3-item task (t₍₈₄₎ = 7.19, p < 0.001, Cohen’s d = 1.55). (C) Saccade amplitude during encoding. A marginally significant difference was observed between the two tasks (t₍₈₄₎ = 1.92, p = 0.06, Cohen’s d = 0.42). (D) Fixation duration during encoding. The two task showed no significant difference (t₍₈₄₎ = 0.91, p = 0.36, Cohen’s d = 0.20). (E–H) similar to (ABCD), but for the maintenance period. The 4-item task showed a greater number of saccades (E; t₍₈₄₎ = 2.58, p = 0.01, Cohen’s d = 0.56) and fixations (F; t₍₈₄₎ = 3.34, p = 0.001, Cohen’s d = 0.72) than 3-item task. No significant differences were observed in saccade amplitude (G; t₍₈₄₎ = 1.42, p = 0.16, Cohen’s d = 0.31) or fixation duration (H; t₍₈₄₎ = 0.27, p = 0.79, Cohen’s d = 0.06). (I) Fixation-sequence analysis in the 3-item task. Left: fixation order showed evidence of a 1 → 2 → 3 sequence relative to chance level (36) during encoding (yellow, t₍₄₅₎ = 2.76, p = 0.008, Cohen’s d = 0.41), but not during maintenance (purple). Right: during encoding, this sequential fixation pattern was present in high-performing participants (t₍₂₂₎ = 3.40, p = 0.003, Cohen’s d = 0.71), but not in low-performing participants (t₍₂₂₎ = 0.28, p = 0.78, Cohen’s d = 0.06). (J) Same as (I), but for 4-item task. The 1 → 2 → 3 → 4 sequence pattern wasn’t significantly expressed at group level (chance level: 18, t₍₃₉₎ = 1.51, p = 0.14, Cohen’s d = 0.24), but significantly expressed in high-performing participants (high-performing: t₍₁₉₎ = 2.31, p = 0.03, Cohen’s d = 0.52; low-performing: t₍₁₉₎ = –1.39, p = 0.18, Cohen’s d = –0.31). Error bars indicate SEM. Dots represent individual participants. * p < 0.05. Note that no reliable fixation sequence was detected during maintenance, likely because many trials contained few overt saccades and fixations. The corresponding data are publicly available on OSF: https://osf.io/kasr2/.

https://doi.org/10.1371/journal.pbio.3004047.s007

(TIF)

S8 Fig. Gaze position during encoding and maintenance in the central-fixation control experiment.

(A) Gaze position time courses along the horizontal (upper panel) and vertical (lower panel) axes during the encoding for the six participants who were excluded because they did not sufficiently maintain central fixation. (B) Gaze position time courses along the horizontal (upper panel) and vertical (lower panel) axes during the encoding period for the remaining 31 participants included in the analyses. (C) Heat maps of the overall gaze distribution for the 31 included participants during encoding (left) and maintenance (right), demonstrating successful fixation at the center of the screen. The corresponding data are publicly available on OSF: https://osf.io/kasr2/.

https://doi.org/10.1371/journal.pbio.3004047.s008

(TIF)

S9 Fig. Subsampling Experiment 1 to match the number of trials per configuration in Experiment 2.

To assess whether the larger number of trials per configuration in Experiment 1 contributed to the observed effects, we randomly subsampled 18 of the 36 trials from each spatial configuration, matching the number of trials per configuration in Experiment 2, and repeated the pattern-similarity and sequenceness analyses. (A) Pattern-similarity effect during encoding (t₍₄₅₎ = 7.77, p < 0.001, Cohen’s d = 1.62) and maintenance (t₍₄₅₎ = 2.11, p = 0.04, Cohen’s d = 0.44). (B) Sequenceness during encoding. Left: sequence strength for the 1 → 2 → 3 and 3 → 2 → 1 sequences across all participants. Right: 1 → 2 → 3 sequence strength shown separately for high- and low-performing participants, defined by a median split of memory performance. (C) Same as (B), but for the maintenance period. Overall, the trial-matched dataset showed a pattern consistent with that observed in the full Experiment 1 dataset. The corresponding data are publicly available on OSF: https://osf.io/kasr2/.

https://doi.org/10.1371/journal.pbio.3004047.s009

(TIF)