Is repulsive serial bias in visual perception driven by adaptation mechanisms?

Abstract

Reported perception can exhibit a repulsive bias away from a task-irrelevant prior stimulus. Previous research has suggested that this repulsive serial bias is driven by low-level adaptation, such that the prior stimulus repels the representation of the new stimulus during encoding. To test this account, the present study compared the repulsive serial bias with another perceptual bias that is known to be driven by an adaptation mechanism (e.g., the tilt aftereffect). We measured the repulsive serial bias using a common location delayed estimation task and the adaptation-driven bias using a location estimation task with an inducer stimulus. We found that, although both repulsive serial bias and adaptation-driven bias were evident, the two biases were not correlated. In addition, only the repulsive serial bias was associated with a response time effect, where responses were slower when the bias was stronger. Moreover, mouse-tracking data for the repulsive serial bias exhibited a pattern that started with a stronger repulsion and ended with smaller repulsion, which cannot be explained by an adaptation mechanism alone. Taken together, our findings suggest that repulsive serial bias in continuous estimation tasks involves post-perceptual decisional processes that are not present in the adaptation-driven bias.

Introduction

Our ongoing perceptual experiences are shaped not only by the stimulus currently in view but also by stimuli seen moments earlier. For example, in orientation estimation tasks, the reported orientation of a stimulus can be biased toward the stimulus from the previous trial, even when that previous stimulus is completely irrelevant to the task at hand (Fischer & Whitney, 2014). This attractive serial dependence is a general phenomenon observed across various types of visual stimuli, including location (Bliss, Sun, & D'Esposito, 2017), color (Barbosa & Compte, 2020), and faces (Liberman, Fischer, & Whitney, 2014). These findings suggest that serial dependence reflects a mechanism by which the visual system achieves perceptual stability over time by integrating recent perceptual history into the processing of new stimuli (Manassi, Murai, & Whitney, 2023).

Although studies on serial dependence have shown that stimulus history biases perception toward prior stimuli, recent research has found that perceptual reports can also exhibit repulsive bias away from the prior stimulus (Akselberg et al., 2025; Bae & Luck, 2019; Bae & Luck, 2020; Bansal et al., 2023; Bliss et al., 2017; Chen & Bae, 2024; Fritsche, Mostert, & de Lange, 2017; Stein et al., 2020). For example, Bae & Luck (2020) found that reports of motion direction were biased away from the prior motion direction. This repulsive serial bias has also been observed in reports of stimulus orientation (Bae & Luck, 2019) and spatial locations (Bliss et al., 2017). Such repulsive biases are typically considered adaptation-driven effects that occur during the encoding of a new stimulus (Fritsche, Spaak, & de Lange, 2020; Moon & Kwon, 2022; Pascucci et al., 2019), similar to the classic tilt-aftereffect (Gibson & Radner, 1937).

Consistent with this view, recent neuroimaging studies have found evidence that neural responses to the current stimulus exhibit a repulsive bias away from the prior stimulus during early visual processing. For example, Sheehan and Serences (2022) showed that stimulus information decoded from patterns of BOLD activity in the visual cortex (e.g., V3) was biased away from the prior stimulus (see also, Luo, Zhang, Fang, & Luo, 2025). However, behavioral reports in these studies typically showed attractive serial dependence, leaving it unclear whether repulsive biases in the reports directly reflect early sensory adaptation itself.

Interestingly, recent studies suggest that the repulsive serial bias may not be driven solely by low-level adaptation mechanisms. When Bae and Chen (2025) measured response times in a motion direction estimation task, they found that the prior stimulus not only repelled reports of the current motion direction but also influenced the speed of responses. Moreover, this response time effect was observed only when the repulsive serial bias was present, suggesting that the repulsive serial bias may involve post-perceptual decisional processes rather than purely low-level adaptation. To further investigate this possibility, the present study directly compared the repulsive serial bias with another repulsive bias known to be driven by low-level adaptation mechanisms, such as the tilt-aftereffect (Gibson & Radner, 1937).

In the experiment, participants performed two types of location estimation tasks. In one task, they reported the location of the sample stimulus after a short delay interval (i.e., delayed estimation task) (Figure 1a), which is commonly used in serial bias studies. We predicted a repulsive serial bias from this task following our previous studies with a similar task design (Bae & Luck, 2019). In another task, participants reported the location of the sample stimulus, which was presented after an inducer stimulus (i.e., inducer task) (Figure 1b). In the analyses, we examined whether the repulsive serial bias from the delayed estimation task is correlated with the adaptation bias from the inducer task. We predicted that, if the repulsive serial bias reflects adaptation to the previous-trial stimulus, then the repulsive serial bias from the delayed estimation task should be positively correlated with the adaptation bias from the inducer task.

Figure 1.

Location delayed estimation task (a) and location estimation task with an inducer (b). In both tasks, participants reproduced the location of the sample stimulus (i.e., black dot) by adjusting the location of the test object using a computer mouse. The test object appeared only after participants moved the mouse to make a report. Response time (RT) was measured by the time difference between the onset of the response cue and the final mouse click. During the mouse report, we measured the mouse trajectory to examine how the response unfolded over time. (c) Prediction of response time. If the bias reflects only the distortions in the perceptual representations themselves, then the speed of response should be independent of the location differences (black). If, however, the bias involves decisional processes, then the speed of response should vary with location differences (cyan). For example, trials with small location differences should exhibit slower RT because a similar location would more strongly interfere with the decision. (d) Predictions of mouse trajectories. Participants may move the mouse directly to the final click location (black) or they may adjust the mouse over time (cyan). Example shown here reflects repulsive serial bias where the report of the current trial (R_N) to the current-trial stimulus (T_N) is biased away from the previous-trial stimulus (T_N-1). The right panel shows the corresponding response trajectories with the target location aligned at 0° on the y-axis.

However, the presence of correlation may not necessarily imply that the two biases were driven by common mechanisms. We therefore scrutinized additional variables—the speed of response and the trajectory of mouse report—for the two biases, which are known to vary with decision-related processes (Resulaj, Kiani, Wolpert, & Shadlen, 2009). For the response time, we predicted that if the repulsive serial bias involves post-perceptual decisional processes that compare the current-trial stimulus with the previous-trial stimulus, then the speed of response may vary with the similarity between the previous and the current stimuli. For example, if the previous stimulus interferes with the decision for a new stimulus, then the speed of response might be slower when the two stimuli are more similar (Bae & Chen, 2025). However, if repulsive serial bias is driven by low-level adaptation mechanisms that directly alter the stimulus representation without the involvement of post-perceptual decisional processes, then the response time should not vary with the similarity between the two consecutive stimuli. We predicted that this should be the case for the repulsive bias from the inducer task as the inducer stimulus would directly alter the stimulus representation via low-level adaptation mechanisms.

For the mouse trajectory, we predicted that if the repulsive serial bias involves decisional processes, then the trajectory may exhibit a pattern that reflects systematic response adjustments. For example, mouse reports may start with a larger repulsive bias away from the previous stimulus and end with a smaller bias (Figure 1d, blue) (Chen & Bae, 2024; Chen & Bae, 2025). However, if the repulsive serial bias reflects the biased representation itself and does not involve post-perceptual decisional processes, then the mouse report should exhibit a pattern that does not show systematic response adjustments. For example, the mouse reports would start from the central fixation and directly go to the to-be-reported location, exhibiting a straight-line-like pattern (Figure 1d, black) (Chen & Bae, 2024). We predicted that this should be the case for the repulsive bias from the inducer task.

To foreshadow our main results, we found that the repulsive serial bias from the delayed estimation task was not correlated with the adaptation bias from the inducer task. The repulsive serial bias was accompanied by both a response time effect and a trajectory effect: reporting the location was slower when the current-trial stimulus was similar to the previous-trial stimulus, and mouse responses began with a larger repulsive bias that was reduced to a smaller bias by the end of the report. Critically, individuals who exhibited a stronger repulsive serial bias also showed stronger response time and trajectory effects. In contrast, the adaptation bias was not associated with either effect, consistent with our predictions. These results suggest that repulsive serial bias and adaptation bias are driven by different mechanisms: repulsive serial bias involves post-perceptual decision processes, whereas adaptation bias reflects distortions in the representation itself.

Method

Participants

A prior study using a similar task design reported a large effect size for a repulsive serial bias effect (Cohen's d > 1) (Bae & Luck, 2019). A power analysis determined that a minimum of 15 samples would be necessary to detect a large effect (Cohen's d = 0.8) with a statistical power of 0.8. The current study recruited a total of 40 participants (21 females; 19 males) between the ages of 18 and 30 with normal or corrected-to-normal visual acuity. All participants provided written informed consent prior to participating in the experiment and received course credit upon completion of the study. The study was approved by the Arizona State University Institutional Review Board in accordance with the Declaration of Helsinki (World Medical Organization, 2013).

Apparatus, stimuli, and procedure

The stimuli were generated using the Psychtoolbox-3 (Brainard, 1997) library and MATLAB (MathWorks, Inc., Natick, MA, USA) and were presented on a 60 Hz LCD monitor at a viewing distance of approximately 70 cm. A black fixation dot was continuously present in the center of the display except during the intertrial interval.

Each participant completed two location estimation tasks—a location delayed estimation task (Figure 1a) and a location estimation task with an inducer (Figure 1b)—in separate blocks of trials with the order being counterbalanced. The delayed estimation task was designed to produce repulsive serial bias. Each trial started with a 500-ms presentation of the central fixation dot followed by a target stimulus (a black dot, 1.64° of visual angle in diameter) for 500 ms. The location of the target stimulus was randomly sampled from a set of 12 discrete locations (0° to 360° in steps of 30°). Participants were asked to remember the location of the stimulus as precisely as possible. After a 1000-ms delay, a response cue (a black ring with a radius of 4°) appeared at the center of the screen with the mouse pointer to indicate the start of the response period. Many prior studies presented a response object with a random value at the start of the response period (e.g., Fischer & Whitney, 2014). Because the random response object can impact response time and mouse trajectory measures, we did not present the response object at the start of the response period (Figure 1a). As soon as participants moved the mouse pointer for a report, the response object (black dot) appeared at the corresponding angular position on the response ring. The angular position of the response object continuously followed the mouse pointer. Participants continuously adjusted the location of the response object and clicked the mouse button to submit the response. The mouse pointer was visible during the movement. To ensure that participants precisely adjusted the location of the response, we had participants move the mouse pointer outside of the response ring. The next trial started after a 500-ms intertrial interval. Each participant completed 16 practice trials and then completed 288 main experimental trials (24 trials for each target location, randomly intermixed) which were presented in three blocks of 96 trials.

The location estimation with an inducer task was designed to induce repulsive low-level adaptation bias away from the inducer. This task was identical to the delayed estimation task except that (1) an inducer stimulus (a white dot, 1.64° of visual angle in diameter) was presented for 2000 ms before the target stimulus (a black dot) and (2) no delay between the target stimulus and the report was introduced to minimize possible impacts of post-perceptual processing on the bias effect. (Figure 1b). A 2000-ms duration for the inducer stimulus should be sufficient to induce low-level adaptation effect given that adaptation occurs after exposure to a stimulus as brief as 50 ms (Bonds, 1991; Felsen et al., 2002; Harris & Calvert, 1989). The locations for the inducer and target stimulus were randomly sampled from a set of 12 discrete locations (0° to 360° in steps of 30°). Each participant completed 16 practice trials and then 288 main experiment trials (12 inducer locations × 12 target locations × 2 trials) which were presented in three blocks of 96 trials.

In both experiments, we recorded response time (RT), the trajectory of the mouse report, and the final response. RT was measured by the time difference between the onset of the response cue and the mouse click. Mouse trajectory was measured by the angular deviation of the response object for every 20 ms during the entire duration of the mouse report. The final response was measured by the angular position of the response object at the time of the mouse click (which corresponds to the final position of the mouse trajectory).

Analyses

Before conducting our main analyses, we excluded trials based on trial exclusion criteria used in our previous studies (Chen & Bae, 2024; Bae & Chen, 2025). Trials with an absolute response error (i.e., reported location minus target location) exceeding 60° were excluded from the analyses (1% of total trials in the delayed estimation task; 0.5% of total trials in the inducer task). Additionally, trials with response time faster than 300 ms or slower than five seconds were excluded (1% of total trials in the delayed estimation task; 0.5% of total trials in the inducer task). For the analyses of serial bias, the first trial of each block was necessarily excluded (three trials total).

Serial bias and adaptation bias analyses

The goal of this analysis is to investigate whether the location reports were biased in relation to the previous-trial location in the delayed estimation task (i.e., serial bias) and also to the inducer stimulus in the inducer task (i.e., adaptation bias). To that end, we first subtracted the grand mean of response errors from individual trials for each observer to compensate for general bias in the location reports (Bliss et al., 2017). We then coded the sign of response errors relative to the prior stimulus and to the inducer. A positive sign was given to response errors deviated away from the prior stimulus or from the inducer, and a negative sign was given to response errors deviated toward it (Bae & Luck, 2019). Therefore a positive response error in our analyses indicates repulsive serial bias or repulsive adaptation bias. The sign was randomly given to trials with 0° and 180° location differences because the direction of the bias was undefined for those trials.

Because stimulus locations were selected from a set of 12 discrete values, each trial corresponded to one of seven absolute location differences (i.e., combining positive and negative differences): 0°, 30°, 60°, 90°, 120°, 150°, or 180°. After trial exclusion (see above), the bins with location differences greater than 0° and less than 180° contained, on average, 46 trials in the delayed estimation task and 48 trials in the inducer task. The 0° and 180° bins contained half as many trials, as the sign of location differences cannot be defined for these cases. Following our previous studies (Bae, 2024; Bae & Luck, 2017; Bae & Luck, 2019), we quantified the magnitude of the serial bias as the average signed response error for trials with small location differences (0° < Δ < 90°; >90 trials per participant) and tested this mean against zero using a one-sample t-test.

Response time (RT) analysis

RT was measured by the time difference between the onset of the response cue, and the final mouse click. Median RT was calculated for each of the seven absolute location difference bins in both tasks. The RT effect was assessed by comparing the mean of the median RT for the trials with small location differences (Δ < 90°) with the mean of the median RT for the trials with large location differences (Δ > 90°). This was statistically tested using paired t-tests.

Mouse trajectory analysis

To investigate how the response unfolded over time during the report, we recorded the angular position of the response object every 20 ms during the mouse report. For each time point in the trajectory, we calculated response errors the same way as in the final response bias analysis and gave a positive or negative sign depending on the relative position of response errors to the previous-trial target or the inducer (e.g., positive sign = repulsive bias). Because the length of the mouse trajectory varied across trials, we computed the average mouse trajectory twice, one by time-locking the trajectories to the initiation of the mouse report to assess how the mouse report started, and one by time-locking the trajectories to the mouse click to assess how the mouse report ended (Chen & Bae, 2024; Chen & Bae, 2025).

The statistical analyses for the mouse trajectory tested whether the trajectories were curved or flat when the bias occurred. To do so, we compared the average of the initial 200 ms from the trajectory time-locked to the mouse response initiation with the average of the final 200 ms from the trajectory time-locked to the mouse click. These parameters for quantifying mouse trajectory were determined based on previous studies (Chen & Bae, 2024; Chen & Bae, 2025). If the mouse trajectory was relatively flat, there would be little to no significant difference between the initial and final 200 ms. However, if the mouse trajectory was curved, there would be a significant difference between them. We tested these possibilities using a paired t-test.

Data availability

The data are publicly available online at https://osf.io/srxgp/.

Results

Repulsive serial bias and adaptation bias

Figure 2a shows serial bias from the delayed estimation task and adaptation bias from the inducer task. Location reports were biased away from the previous-trial stimulus location (i.e., positive error) when the two locations were similar. And they were biased away from the inducer stimulus (i.e., positive error) when the inducer was similar to the sample location. To test these effects statistically, we computed the average response error for the trials with small location differences (0° < Δ < 90°) and compared it against zero. One-sample t-tests confirmed that both the repulsive serial bias (M = 0.91°, t(39) = 6.33, p < 0.001, 95% confidence interval [CI] = [0.62°, 1.21°], Cohen's d = 1) and adaptation bias (M = 0.53°, t(39) = 3.08, p = 0.007, 95% CI = [0.18°, 0.87°], Cohen's d = 0.49) were statistically significant. However, the two biases were not significantly correlated (t(38) = −0.74, p = 0.46, 95% CI = [−0.415°, 0.200°], r = −0.12, BF₀₁ = 2.24) (Figure 2b).

Figure 2.

(a) Serial bias (black) from the delayed estimation task and adaptation bias (red) from the inducer task. Positive value indicates repulsive bias. Error bar: ±1 SE. (b) Correlation between serial bias and adaptation bias. The magnitude of the bias effect was quantified by the average of the bias for small location differences (0° < ∆ < 90°). Each circle represents an individual observer. The solid line represents the best-fit linear model.

RT

Figure 3a shows the mean of median RTs as a function of location differences between the current- and the previous-trial stimuli for the serial bias, and between the target and the inducer stimuli for the adaptation bias. Location reports tended to be slower when the location for the previous-trial stimulus was similar to the location for the current-trial stimulus. However, response time was comparable irrespective of the location difference between the target and the inducer stimuli. To test this effect statistically, we computed the RT effect by taking the difference between the RTs for small location differences (Δ < 90°) and the RTs for large location differences (Δ > 90°). A paired t-test confirmed a significant RT effect for the repulsive serial bias (M = 70 ms, t(39) = 4.32, p < 0.001, 95% CI = [35.57 ms, 95.54 ms], Cohen's d = 0.68). The RT effect for adaptation bias was not significant (M = −1.25 ms, t(39) = −0.11, p = 0.92, 95% CI = [−25.06 ms, 22.57 ms], Cohen's d = 0.017, BF₀₁ = 5.83). A paired t-test confirmed a significant difference in the RT effect between the two biases (M = 66.25 ms, t(39) = 3.73, p < 0.001, 95% CI = [30.37 ms, 102.13 ms], Cohen's d = 0.59). In supplementary analysis, we confirmed that the absence of the RT effect for the adaptation bias was not caused by the absence of the delay period in the task (Supplementary Figures S1 and S3).

Figure 3.

(a) Median response time (RT) as a function of location differences for serial bias in the delayed estimation task (black) and for adaptation bias in the inducer task (red). Error bar: ±1 SE. (b) Correlation between the repulsive serial bias and the response time effect (i.e., RT for small location differences minus RT for large location differences). The repulsive serial bias was positively correlated with the RT effect. (c) The correlation between the adaptation bias and the RT effect. The adaptation bias was not correlated with the RT effect. Each circle represents an individual observer. The solid lines represent the best-fit linear models.

If the RT effect for the repulsive serial bias is not a mere artifact of the task and truly related to the bias effect, then the RT effect should be tightly linked to the bias effect. To test this prediction, we conducted a correlation analysis and found that the RT effect was positively correlated with the repulsive serial bias: individuals who exhibited a stronger RT effect also exhibited a stronger repulsive serial bias, (t(38) = 2.14, p = 0.04, 95% CI = [0.02, 0.58], r = 0.33) (Figure 3b). However, the RT effect was not significantly correlated with adaptation bias (t(38) = −0.11, p = 0.917, 95% CI = [−0.33, 0.30], r = −0.02, BF₀₁ = 2.83) (Figure 3c). For the sake of completeness, we report the correlation between the RT effect and the bias at large stimulus differences in the Supplementary material.

Mouse trajectory

Figure 4a shows the mouse trajectory for the repulsive serial bias (black) and adaptation bias (red). For the serial bias, location reports started with a strong repulsive bias and ended with a smaller repulsive bias. A similar pattern of response trajectory was observed for the adaptation bias. However, the magnitude of the initial repulsion for the adaptation bias was much smaller compared to the initial repulsion for the repulsive serial bias. To test these trajectory effects statistically, we compared the average of the first 200 ms and the average of the last 200 ms using a paired t-test. The trajectory effect was significant for both repulsive serial bias (M = 11.91°, t(39) = 7.88, p < 0.001, 95% CI = [8.85°, 14.97°], Cohen's d = 1.25) and adaptation bias (M = 1.76°, t(39) = 2.87, p = 0.007, 95% CI = [0.52°, 3.00°], Cohen's d = 0.45). However, the magnitude of the initial repulsion for serial bias was 5.7 times greater compared to the initial repulsion for adaptation bias (M = 10.55°, t(39) = 6.90, p < 0.001, 95% CI = [7.46°, 13.65°], Cohen's d = 1.09). In a supplementary analysis, we confirmed that the weak initial repulsion for the adaptation bias was not caused by the absence of the delay period in the task (Supplementary Figures S1 and S3).

Figure 4.

(a) Mouse trajectories averaged across the trials with small location differences (0° < Δ < 90°) for the repulsive serial bias (black) from the delayed estimation task and the adaptation bias (red) from the inducer task. The left panel shows the trajectories time-locked to the initiation of the mouse response and the right panel shows the trajectories time-locked to the mouse click. Positive errors indicate bias away from the prior stimulus (black) or from the inducer (red). The shaded areas represent ±1 SE. (b) Correlation between the initial bias (average of the first 200 ms) and the final bias (average of the last 200 ms) in the trajectory for the serial bias. (c) Correlation between the initial bias (average of the first 200 ms) and the final bias (average of the last 200 ms) in the trajectory for the adaptation bias. The solid lines represent the best-fit linear models.

To examine whether the initial repulsion in the trajectory is independent of the final bias effects, we tested whether the magnitude of the initial repulsion was correlated with the strength of the bias at the end of the report. We found that individuals exhibiting a stronger initial repulsion also showed a stronger repulsive serial bias at the end of the report (t(38) = 2.40, p = 0.02, 95% CI = [0.058, 0.61], r = 0.36). This result indicates that the final repulsive serial bias reflects the strength of the initial repulsion. In a supplementary analysis, we found that the magnitude of the reduction in initial repulsion (i.e., initial repulsion minus final bias) also depended on the strength of the initial repulsion (see Supplementary material). In contrast, when we assessed the relationship between initial repulsion and the adaptation bias measured in the inducer task, we found no correlation between initial repulsion and final bias (t(38) = −0.02, p = 0.88, 95% CI = [−0.34, 0.29], r = −0.03, BF₀₁ = 2.81).

Discussion

Past perceptual experience can repel future perceptual decisions away from it, even when the experience is no longer relevant to the current goal of behavior. Prevailing theories have proposed that this repulsive serial bias is driven by low-level adaptation mechanisms that alter the stimulus representation itself at the perceptual stage of processing. However, there is a shortage of evidence that the repulsive serial bias directly reflects the adaptation process. Here, we tested this proposal by directly comparing the repulsive serial bias with another perceptual bias that is known to be driven by a low-level adaptation mechanism. In addition to the final bias, we also assessed other aspects of behavior that have been overlooked in the literature, response time and mouse trajectory, to better understand how the bias unfolded over time. Our results showed that, although both repulsive serial bias and adaptation-driven bias were evident, they were not correlated. Moreover, only the repulsive serial bias was associated with specific patterns in response time and mouse trajectory. These results provide converging evidence that an adaptation mechanism alone is not sufficient and post-perceptual decision processes are necessary to fully account for the repulsive serial bias in continuous estimation tasks.

Prior stimuli modulate the speed of decision for a new stimulus

Although most of the research on serial bias focuses on how the stimulus history biases perceptual reports, recent studies found that the stimulus history also modulates the speed of response for a new stimulus. Bae and Chen (2025) showed that the speed of motion direction estimation was slower when the motion direction in the previous trial was similar to the motion direction in the current trial. Importantly, this RT effect was present only when the repulsive serial bias was present, and the RT effect was correlated with the magnitude of the repulsive serial bias. Consistent with these findings, the present study also shows that the RT effect occurs for the repulsive serial bias with location estimation, that the effect is correlated with the final bias, and that it does not occur for the final bias that is primarily driven by low-level adaptation. These results provide converging evidence that the RT effect was driven by the mechanism for the repulsive serial bias and suggest that post-perceptual mechanisms that determine the speed of response should be incorporated into the theories in serial bias.

Initial repulsion in the trajectory reflects adaptation-driven bias

The curved pattern of the response trajectory provides additional evidence that repulsive serial bias involves post-perceptual decision processes. Specifically, the results suggest that at least two mechanisms are required to fully explain the pattern of the response trajectory. First, an adaptation mechanism repels the representation of the new stimulus away from the previous stimulus during early processing (i.e., the initial repulsion in the trajectory). Second, decisional mechanisms counteract this repulsion, leading to a smaller but still significant bias at the end of the trajectory. Our findings indicate that these mechanisms operate at different processing stages—early adaptation and post-perceptual decision—and that post-perceptual decision processes depend on the strength of the adaptation: the magnitude of the final repulsive bias was predictable from the magnitude of the initial repulsion in the trajectory (Figure 4b).

The present study does not provide direct evidence that the initial repulsion in the trajectory was driven by adaptation mechanisms. However, the observed pattern of initial repulsion is consistent with predictions based on adaptation. First, the magnitude of the initial repulsion was stronger when two consecutive stimuli were similar (see Supplementary Figure S2), reflecting the feature selectivity of adaptation. Second, a recent study showed that the magnitude of initial repulsion decreased when stimulus locations varied randomly across trials and when the delay interval increased (Chen & Bae, 2025), reflecting the spatial and temporal dependence of adaptation. Third, the magnitude of initial repulsion in the trajectory (Figure 4a) was comparable to the magnitude of the repulsive adaptation bias estimated from neural activities in early visual cortex (Luo et al., 2025; Sheehan & Serences, 2022). Finally, a recent study showed that initial repulsion in the trajectory was absent for perceptual biases not driven by adaptation (e.g., cardinal bias) (Chen & Bae, 2024). Together, these findings provide converging evidence that the initial repulsion in the trajectory is driven by adaptation mechanisms.

Our analyses related the current-trial stimulus to the previous-trial stimulus, and therefore our results suggest that the initial repulsion was driven by the adaptation to the previous stimulus, consistent with previous studies (Fritsche et al., 2017; Sadil, Cowell, & Huber, 2023). However, it is also possible that the initial repulsion was driven by the previous decision. In line with this interpretation, past research has shown that the previous-trial stimulus did not lead to repulsive serial bias at all if it was not reported, even if it was perceived and encoded into working memory (Bae & Luck, 2020). In addition, a recent neural decoding study provided converging evidence that the previous-trial decision serves as an inducer of the adaptation effect for the subsequent stimulus (Luo et al., 2025). From this perspective, the weaker adaptation observed in the inducer task may arise because no explicit decision was required for the inducer stimulus. However, the present study was not designed to investigate the exact source of the initial repulsion; therefore, we do not make a strong conclusion about the role of the previous-trial stimulus and decision in the initial repulsion. Future studies should investigate this issue by dissociating the stimulus and response in a task.

Common mechanisms may underlie both repulsive and attractive serial biases

The present study examined the nature of repulsive serial bias in relation to adaptation-driven repulsive effects. Therefore we do not draw conclusions regarding the nature of attractive serial dependence. However, it is worth noting that previous evidence indicates that attractive serial dependence is also linked to decisional dynamics similar to those observed here: response times were slower for trials with small stimulus differences, and mouse trajectories initially showed a stronger repulsive bias even when the final response exhibited attractive serial dependence (Chen & Bae, 2024; Chen & Bae, 2025). These findings suggest that attractive serial dependence, like repulsive bias, involves post-perceptual decisional processes rather than reflecting a purely perceptual bias.

The presence of common decisional dynamics in both repulsive and attractive biases suggests that the difference in bias direction reflects a quantitative, rather than qualitative, difference in decisional processes. Specifically, the visual system must account for variations in visual input that occur across time and space to interact effectively with objects and scenes in the environment. This sensitivity is achieved through adaptation mechanisms, which adjust perceptual representations based on recently encountered stimuli. However, because adaptation can substantially bias perception away from recent stimuli, mechanisms that mitigate this bias are necessary to achieve perceptual stability over time. We propose that the visual system balances stability and sensitivity by differentially correcting adaptation-induced biases during post-perceptual processing, depending on contextual demands (Chen & Bae, 2025). For example, when sensory evidence is less reliable due to extended working-memory delay (e.g., Bliss et al., 2017; Chen & Bae, 2025), a corrective adjustment may be applied to counteract and overcome adaptation-induced repulsive biases, thereby minimizing errors and stabilizing perception over time. Conversely, when sensory evidence is highly reliable, a corrective adjustment may instead attenuate large adaptation-induced repulsive biases to support greater decisional sensitivity while reducing errors in the reports. Future studies investigating this possibility could provide significant new insights into how stimulus history shapes our perceptual experiences across different task contexts.