Reactivation of previous decisions repulsively biases sensory encoding but attractively biases decision-making

Perception is shaped not only by current sensory inputs but also by temporal regularities in the environment, producing history-dependent effects such as serial dependence, where current perception is biased toward previous trials. Serial dependence has been reported across a wide range of visual features (orientation, numerosity, position, facial attractiveness) and modalities. While this can reduce perceptual fidelity, it may optimize stability and efficiency when the world is stable over short timescales. Two main accounts debate the stage at which serial dependence arises. The continuity field account posits a perceptual-level integration of nearby stimuli, yielding attractive biases. In contrast, a post-perceptual account proposes a two-stage process: efficient sensory coding producing repulsion and a subsequent Bayesian inference producing attraction. Neural evidence has been mixed: some studies observe attractive shifts in visual cortex consistent with the continuity field, whereas others show repulsive neural biases despite attractive behavior, and causal evidence implicates higher-order regions (PFC, PPC) in attractive bias. Serial dependence implies that past-trial memory traces influence current processing. Prior work shows transient reactivation of past information before trial onset correlating with bias behavior, and feature-specific reactivation of past-trial information triggered by current events. However, memory reactivation alone cannot explain serial bias; interactions between reactivated past information and current processing likely occur across stages and regions. Theoretical and empirical work suggests the brain organizes distinct information into (near-)orthogonal subspaces to reduce interference. The authors hypothesize that non-orthogonal alignment between past reactivations and present representations permits interaction: aligned axes produce attractive bias; flipped axes produce repulsive bias. The study uses delayed-response tasks and time-resolved EEG/MEG to examine when, where, and how serial dependence emerges, explicitly separating encoding and decision-making. It also dissociates stimuli from reports to identify which past information (stimulus vs. report) influences current perception and choice. Behaviorally, participants show attractive bias toward previously reported locations. Neurally, past reports reactivate and interact with current processing in a two-stage manner: repulsive during encoding in sensory cortex and attractive during decision-making in prefrontal cortex, with the late stage modulated by task relevance and predicting behavior.

The paper reviews extensive evidence for serial dependence across visual and other modalities and contrasts two accounts: (1) a perceptual continuity field integrating nearby inputs to yield attraction; (2) a two-stage account combining efficient coding (repulsion) with Bayesian inference (attraction). Mixed neural evidence includes attractive patterns in V1/V2 (supporting perceptual-level attraction) and repulsive neural biases alongside attractive behavior, as well as causal involvement of PFC/PPC in attractive serial dependence (supporting a post-perceptual locus). The literature on activity-silent working memory and reactivation shows past-trial features can be reactivated by current events and influence ongoing processing. Theoretical frameworks on subspace organization propose orthogonalization to minimize interference; recent studies in monkeys and humans demonstrate rotation/partition of representations to separate sensory and mnemonic content. The authors leverage these insights to predict that non-orthogonal (aligned or flipped) relationships between past and present representational axes determine the direction (attractive vs. repulsive) of serial biases.

Participants and ethics: Experiment 1 (EEG): 31 recruited, 29 analyzed (13 females; 18–27 years, mean 21.6). Experiment 2 (MEG): 27 recruited, 26 analyzed (17 females; 19–26 years, mean 21.6). Informed consent obtained; ethics approval (#2021-10-15), Peking University. Apparatus: Experiment 1 used a 32-inch Display++ LCD (100 Hz, 1920×1080) at 57 cm; Experiment 2 used a 26-inch rear-projection screen (60 Hz, 1024×768) at 75 cm. Gray background; fixation maintained. Tasks:

• Experiment 1 (EEG, delayed-location reproduction): On each trial, a red dot (0.4°) appeared within a 15° diameter circular region for 520 ms (Encoding), followed by a 1020 ms Gaussian-smoothed white-noise mask (Delay). Then a mouse cursor appeared at a random location; participants reproduced the target location by moving/clicking; spacebar confirmed response (no time limit). Inter-trial interval (ITI) 2000 ms. 32 blocks × 50 trials.
• Experiment 2 (MEG, retro-cued 2-AFC): Two colored dots (red/blue) at random locations (500 ms Encoding), a 1000 ms mask (Delay), then a retro-cue (fixation color) indicating which item is target for a location-comparison decision. After 500 ms, two response probes appeared; participants chose which probe was closer to the cued target (keyboard 1/2; no time limit). ITI 2000 ms. 5 blocks × 100 trials. Practice with ≥75% accuracy required. Behavioral analyses:
• Experiment 1: Multiple linear regression of Current Report on Current Stimulus, Current Start (mouse start), and Previous Report for X and Y coordinates, averaged per participant. β3 (Previous Report) indexes serial bias (positive = attraction; negative = repulsion).
• Experiment 2: Due to high accuracy and RT–difficulty coupling, a continuous “weighted choice” was computed: categorical choice coded −1/1 and multiplied by (1 − RT_rescaled). Linear mixed-effects models (identity link; random effects on β) predicted weighted choice from ΔL_Demean for six candidate locations: current target/non-target stimuli, previous target/non-target stimuli, previous chosen/unchosen probes. ΔL = d1 − d2 (distance differences to probes), de-meaned by inherent probe bias via Monte Carlo sampling. Model comparison via AIC against a base model (current target only) using stepwise inclusion. Neural data acquisition and preprocessing:
• EEG: 64-channel actiCAP; FCz reference; 500 Hz sampling; EOG recorded; impedance <15 kΩ. Data epoched (−0.4 to +2.5 s from stimulus), re-referenced to average, baseline-corrected (−0.4 to 0 s), band-pass 2–45 Hz, downsampled to 100 Hz; noisy channels interpolated; ICA to remove ocular/other artifacts.
• MEG: 306 sensors (102 magnetometers, 204 gradiometers; Elekta Neuromag). tSSS denoising (MaxFilter); 1000 Hz sampling; head position controlled (<3 mm between blocks). Preprocessing similar to EEG: band-pass 1–40 Hz, downsample 200 Hz. Structural T1 MRI acquired for source modeling. Neural decoding (trial-wise RSA): For each pair of trials, computed location RDMs (|Loc_i − Loc_j| for X and Y) and neural RDMs based on multivariate patterns within 50 ms windows. For MEG, PCA retained components explaining 99% variance. At each time point, regressed neural RDM on location RDM to obtain representational strength β1t (averaged over X and Y). For decoding of previous-trial variables, current-trial effects were regressed out and residuals analyzed; control analyses showed no spurious decoding of next-trial information. Past–present interaction analysis: At each time point, for each sensor, regressed signals onto location predictors to obtain representational axes (vectors of regression coefficients). Experiment 1 EEG model predictors: current stimulus, current start, previous report. Experiment 2 MEG predictors: current target, current non-target, previous chosen, previous unchosen, current probe1, current probe2 (gradiometers used). Computed cosine similarity between axes for past and present variables: negative (flipped) indicates repulsive interaction; positive (aligned) indicates attractive interaction; zero indicates orthogonality (no interaction). Interactions computed separately for X and Y and averaged; smoothed with a 50 ms window. Statistics: Time-resolved nonparametric sign-permutation with 100,000 permutations and cluster-based correction (p<0.05, two-sided) assessed significance for decoding and interaction time courses. Neuro–behavioral correlations: Experiment 1 across-participant Pearson correlation between behavioral β_Previous Report and neural interaction (averaged within significant clusters) for encoding and decision-making. Experiment 2 within-participant trial-wise regression between inverse RT and trial-level neural interaction (leave-one-out estimation) within encoding/decision clusters; across-participant correlations also examined (reported in supplements). Spatial analyses (MEG): Sensor segregation by anterior (n=156) vs posterior (n=150) sensors. Source localization via LCMV beamforming (5% regularization; BEM forward model; MNE co-registration; 4096 vertices/hemisphere). Sources mapped to 34 Desikan–Killiany regions; decoding and interaction analyses conducted per region; spatial-temporal correction by comparing regional clusters to permuted clusters across all regions.

Behavioral serial dependence:

• Experiment 1 (EEG, continuous 2-D reproduction): Current reports were influenced by current stimulus (mean=0.92, 95% CI [0.90, 0.94], t(28)=84.77, p<0.001, d=15.7), current start (mean=0.0080, 95% CI [0.0044, 0.012], t(28)=4.56, p<0.001, d=0.85), and previous report (mean=0.0039, 95% CI [0.001, 0.0069], t(28)=2.78, p=0.0096, d=0.52), indicating an attractive serial bias toward the previous report.
• Experiment 2 (MEG, retro-cued 2-AFC): Model comparison favored inclusion of previous choice factors over previous stimuli (ΔAIC_best = −35.57 relative to base). In the best LMM, coefficients showed attraction by current target (0.13±0.016, t(25)=42.80, p<0.001, d=8.39) and previous chosen (0.0039±0.0082, t(25)=2.43, p=0.023, d=0.48), and repulsion by current non-target (−0.0052±0.0092, t(25)=−2.91, p=0.0075, d=0.57) and previous unchosen (−0.0038±0.0055, t(25)=−3.53, p=0.0016, d=0.69). This dissociates past choice from past stimulus as the primary driver of attractive bias. Accuracy=0.90±0.033; RT=1.77±0.57 s. Neural reactivation (RSA):
• Experiment 1: Current location decoded during encoding (20–520 ms), delay (0–320 ms, 360–490 ms post-mask), and decision-making (60–760 ms post-cursor). Previous report reactivated during encoding (120–260 ms post-stimulus) and decision-making (230–300 ms, 330–410 ms post-cursor).
• Experiment 2: Current target decoded during encoding (0–500 ms), delay (0–610 ms post-mask), recalling (0–500 ms post-cue), and decision-making (0–1000 ms post-probe). Current non-target also represented during recalling (0–140 ms, 185–500 ms post-cue) and decision-making (0–235 ms, 280–1000 ms post-probe). Previous chosen reactivated in encoding (150–500 ms), delay (140–310 ms post-mask), and decision-making (320–405 ms, 620–765 ms post-probe); previous unchosen reactivated only in early encoding (270–420 ms). Past–present interactions (cosine similarity of representational axes):
• Experiment 1: Two-stage profile with repulsive (negative) interactions during encoding (80–370 ms post-stimulus) and attractive (positive) interactions during decision-making (250–290 ms post-cursor). Neuro–behavioral correlation: late attractive interaction correlated with attractive bias behavior (r(28)=0.37, p=0.045); early repulsion did not (r(28)=−0.063, p=0.75).
• Experiment 2: Early encoding showed repulsive interactions for both task-relevant (target: 170–255, 315–385, 405–460 ms) and task-irrelevant (non-target: 170–265, 290–395, 415–490 ms) conditions. During decision-making, attractive interactions emerged only for task-relevant target (230–400 ms post-probe), not for non-target. Within-participant neuro–behavioral regression showed that late attractive interaction predicted faster responses (β=5.02×10⁻⁴±9.62×10⁻⁴, t(25)=2.66, p=0.014, d=0.52); early repulsive interaction did not (β=8.38×10⁻⁵±7.19×10⁻⁴, t(25)=0.59, p=0.56). Spatial origin (MEG):
• Sensor-level segregation: Posterior sensors dominated early encoding with previous-choice reactivation (150–370, 405–495 ms) and repulsive interactions (175–275, 310–390 ms). Anterior sensors dominated decision-making with past reactivation (110–410, 605–755 ms) and attractive interactions (365–510, 540–635 ms).
• Source-level (Desikan–Killiany regions): Encoding-stage past reactivation in lateral occipital, cuneus, pericalcarine, lingual, parahippocampal; repulsive interaction localized to cuneus (spatio-temporally corrected, p<0.05). Decision-stage past reactivation in medial OFC; attractive interaction in pars orbitalis (inferior frontal gyrus) (spatio-temporally corrected, p<0.05). Additional analyses: Decoding-based reconstruction showed past reports negatively bias reconstructed locations during encoding and positively during decision-making; controls showed no spurious decoding of next-trial variables.

The study resolves a core debate about the locus and mechanism of serial dependence by demonstrating a two-stage process that unfolds across distinct cortical systems. Past choices (goal-based reports) rather than past stimuli form the priors that influence subsequent behavior. These priors are stored in multiple formats and reactivated at specific stages. During sensory encoding in early visual cortex, past reactivation interacts with current representations in a flipped (negative cosine similarity) geometry, consistent with efficient coding and adaptation that repels current representations and may enhance novelty detection. During decision-making in prefrontal cortex (medial OFC and pars orbitalis), past reactivation aligns with current decision variables, consistent with Bayesian integration of prior and evidence, yielding attractive bias and predicting behavior and RT benefits. Task relevance further modulates late interactions: alignment occurs for targets but is orthogonal/non-significant for non-targets, suggesting representational orthogonalization to mitigate distractor interference. The non-orthogonal geometry between past and present representations provides a unifying mechanistic account of both repulsive and attractive components and reconciles prior contradictory neural findings by situating them at different processing stages and regions.

Across two delayed-response experiments with EEG/MEG, the authors delineate a unified operational mechanism for serial dependence: past choices (reports) are reactivated at both encoding and decision-making stages, interacting with current processing in two distinct ways and loci—repulsion in early visual cortex during encoding and attraction in prefrontal cortex during decision-making. The late, task-relevant attractive interaction predicts behavioral bias and RT benefits, whereas the early repulsive interaction reflects automatic adaptation-like efficient coding. By dissociating past stimulus from past choice and separating trial stages, the work clarifies the generation and application of priors in serial dependence and offers a generalizable framework for history-dependent influences across cognition. Future research could causally test stage- and region-specific contributions (e.g., TMS/perturbation), examine the subspace geometry at finer scales and across modalities, and explore how task demands reconfigure representational alignment to balance stability and novelty.

• In Experiment 2, very high accuracy necessitated use of RT-weighted choice and inverse RT as behavioral indices, which may limit comparability to standard accuracy-based measures and could reduce across-participant sensitivity (across-participant neuro–behavioral correlations were nonsignificant; detailed in supplements).
• While MEG source localization implicated specific regions (e.g., cuneus, medial OFC, pars orbitalis), spatial resolution and inverse-model assumptions (e.g., LCMV parameters, head modeling) constrain anatomical specificity.
• The tasks and stimuli were spatial and visual; generalization to other features/modalities, timescales, or more naturalistic settings remains to be directly tested.
• Eye movements were addressed via ICA and stage dissociations, but residual oculomotor confounds cannot be entirely ruled out.
• The study is correlational with respect to neural interactions and behavior; causal evidence for the two-stage mechanism would benefit from perturbation approaches.