Rebuilding structure that arrives out of order
Human experience arrives as a linear stream, but much of what we need to understand, such as the laws of physics or the plot of a novel, is organized as a relational network rather than a temporal chain. The paper asks how the brain recovers that latent structure when the input order is destroyed, and specifically whether the internal model emerges gradually or abruptly after a period of processing.
The authors treat this as a problem of active inference: learners must hold fragmented pieces in working memory and infer the causal links between events that were presented far apart. Prior work had mostly measured the final product of comprehension, through retrospective recall or discrete state transitions, leaving the moment-to-moment trajectory of reconstruction unobserved.
Turning a concept network into a rating task
Participants watched a 12-minute lecture on complex physics concepts, but the video paused after each knowledge event. At each pause, learners first generated a short label for the event as a memory cue, then rated its network distance on a 5-point scale, distinguishing central causes (short path) from peripheral details (long path). The interface displayed all preceding events with their labels and current ratings, so participants could revise earlier judgments as new information arrived.
This design produces a measurable analogue of the expert-defined structure. The authors computed the Pearson correlation between each participant's final rating vector and the expert network-distance levels, and tracked cumulative correlations at each rating stage to see when the learner's model began to match the expert structure. Analyses were restricted to Events 10 to 43 because correlations with fewer than 10 data points are unreliable.
Three conditions separated the role of order: well-structured lectures, ill-structured (scrambled) lectures, and a no-structure baseline. A separate cohort of 73 adults performed the same learning inside a 3.0T MRI scanner, with knowledge tested at pre-test, immediate post-test, a 12-minute delay, and a 24-hour delay.
Disordered input needs about 21 events
The well-structured condition produced immediate and continuous alignment with the expert structure. The ill-structured condition instead showed a temporal latency: accurate structural representation emerged only after the accumulation of approximately 21 events. Before that point, learners were holding fragments without a coherent model.
The period just before structural emergence, Events 18 to 22, functioned as a reconstruction window. During this phase participants made a burst of retrospective updates, with steeper rate-updating slopes and larger update magnitudes. The authors interpret this as a shift from local buffering to global re-evaluation, in which learners retrieve past events and form new links across temporally dispersed concepts. The latency is therefore described as covert computational accumulation rather than passive delay.
The PCC aligns while the model is being revised
To connect behavior to brain activity, the authors used representational similarity analysis. They built neural representational dissimilarity matrices from inter-subject pattern correlation, which isolates shared stimulus-driven processing from individual-specific noise, and correlated those matrices with the expert-defined knowledge structure.
The posterior cingulate cortex showed significant representation of the expert structure only in the ill-structured condition (right PCC r = 0.017, p = 0.030, FDR-corrected), with no significant representation in the well-structured or no-structure conditions. A linear mixed-effects model confirmed a main effect of condition (F(2, 70) = 4.825, p = 0.011), with the ill-structured condition stronger than both the well-structured (Cohen's d = 0.678) and no-structure (d = 0.585) conditions. A leave-one-dyad-out cross-validation replicated this, with bilateral PCC predicting the expert structure in the ill-structured condition.
The effect emerged from Event 18 and tracked the dynamic updating of the learner's internal model. That timing places the PCC in the revision phase, when incoming information must be related to an emerging structure.
The precuneus holds the structure once it stabilizes
The precuneus behaved differently. Robust representations of the expert structure appeared in the bilateral precuneus in the well-structured condition (left r = 0.022, p = 0.015; right r = 0.024, p = 0.010) and also in the ill-structured condition (left r = 0.016, p = 0.030; right r = 0.030, p = 0.010), but not in the no-structure condition. The two structured conditions did not differ significantly from each other (d = 0.001).
The precuneus came online later, from Event 30, after the ambiguity had been resolved. The authors read this as maintenance of the reconstructed model rather than the work of building it. One asymmetry qualifies the picture: in the left precuneus, predictive accuracy was higher in the well-structured than the ill-structured condition (d = 1.163), while the right precuneus showed no such difference, suggesting the representation is more directly accessible when the input was already ordered.
PCC–precuneus coupling predicts immediate comprehension only
Inter-subject functional connectivity between the left PCC and left precuneus was marginally higher in the ill-structured condition than in the well-structured condition (t(70) = 2.378, p = 0.060, d = 1.612), and other region pairs showed no such difference. The authors describe this pathway as a transient functional bridge between updating and maintenance.
That bridge predicted learning success, but only for immediate comprehension of disordered input. Connectivity strength correlated with immediate gains (Test 1 minus Test 0) exclusively in the ill-structured condition, across all four PCC–precuneus pairings (r = 0.412 to 0.483, p = 0.024 to 0.028, FDR-corrected). No significant correlations appeared in the other two conditions at Test 1, and none appeared at the 12-minute or 24-hour assessments in any condition. Coupling with control regions, including early auditory cortex, medial prefrontal cortex, and hippocampus, did not reproduce the pattern.
Immediate parity does not survive 24 hours
Behaviorally, reconstruction was effective in the short run. Immediately after learning, both the ill-structured (d = 2.596) and well-structured (d = 3.259) conditions showed substantial gains over the no-structure baseline, and the ill-structured group was statistically comparable to the well-structured group at both the immediate test (p = 0.148) and the 12-minute delay (p = 0.757).
The 24-hour test separated them. The well-structured group held stable performance between the 12-minute and 24-hour tests (p = 1.000, d = 0.185), whereas the ill-structured group declined significantly relative to its own 12-minute performance (p = 0.046, d = −0.403) and scored lower than the well-structured group (p = 0.026, d = −1.012). The ill-structured group still outperformed the no-structure control (d = 2.017).
The authors suggest that the reconstruction demands may constrain the quality of memory representations formed during encoding. The 24-hour window cannot show whether the gap would widen or close over longer intervals, and the paper explicitly leaves that question open.
A division of labor inside the posterior medial cortex
The paper's central claim is that the posterior medial cortex is not a single integration hub. The PCC carries the load while structure is being inferred and revised, and the precuneus takes over once the model is coherent. Static measures would blend these stages; only the time-resolved trajectory separates them.
For readers weighing disordered or fragmented material, the practical split follows the evidence. The immediate comprehension benefit is real and tied to PCC–precuneus coupling, but the same coupling says nothing about 24-hour retention, and the ill-structured group's delayed drop is the specific cost to watch. The paper's own comparison is between input orders, not between learners, so the relevant threshold is the accumulation point: structure from scrambled input appeared only after roughly 21 events, and the reconstruction window just before it was where retrospective updating concentrated.
Sources used in this answer
Neural dynamics of reconstructing knowledge structure from temporally disordered inputs
Knowledge reconstruction from scrambled lectures followed a delayed two-stage trajectory: the PCC aligned with structure during updating, the precuneus sustained it after reorganization, and their coupling predicted immediate but not 24-hour retention.
