Skip to main content
NCBI home page
NIHPA Author Manuscripts logoLink to NIHPA Author Manuscripts
. Author manuscript; available in PMC: 2026 Apr 22.
Published in final edited form as: J Gerontol B Psychol Sci Soc Sci. 2026 Apr 7;81(5):gbag038. doi: 10.1093/geronb/gbag038

Multi-step inference across the human lifespan can be improved with individualized memory interventions

Sharon M Noh 1, Keiland W Cooper 2, Shuheng Guo 1, Dale Zhou 1,2, Craig E L Stark 2,3, Aaron M Bornstein 1,3,*
PMCID: PMC13099142  NIHMSID: NIHMS2161888  PMID: 41834134

Abstract

Objectives.

Effective goal-directed decision-making relies on memory and planning–processes that are known to decline with age. We tested the hypothesis that these declines stem from a common mechanism by focusing on mnemonic discrimination, a measure of memory precision that shows unique vulnerability to age-related decline.

Methods.

We used a latent learning task that measures the ability to learn and make judgments about associations among interconnected stimuli, assessing performance across the adult lifespan. This task allows us to measure multi-step inference judgments that reflect how individuals organize relational structure, previously shown to capture the internal model-construction processes that support model-based planning. In Experiment 1, we examined relationships between judgment performance and memory precision. In Experiment 2, we tested whether a “blocked” learning schedule designed to reduce memory interference by separating overlapping objects could improve performance for individuals with weaker memory abilities. Across the lifespan, both young and older adults showed evidence of successful latent learning and inference, but variability in judgment performance was explained by mnemonic discrimination ability. In Experiment 2, mnemonic discrimination interacted with training condition: intermixed training benefited those with high memory precision, whereas blocked training benefited those with low memory precision. We also implemented artificial neural network simulations, which reproduced these qualitative patterns.

Discussion.

These findings suggest that age-related declines in complex judgments stem from declines in mnemonic discrimination and demonstrate that individualized, memory-based training interventions can improve learning and reasoning processes that support goal-directed planning, offering a promising approach to preserving decision-making abilities across the lifespan.

Keywords: Individual differences, Latent learning, Memory precision, Training intervention, Goal-directed planning

Introduction

Humans across the lifespan must often make decisions with lasting consequences for themselves and others. Increasingly, older adults occupy positions of power at the highest levels of government and major corporations, making their decision-making abilities particularly consequential for constituents and employees (Akhtar, 2019; Hall & Hickey, 2022; Ingraham, 2014; Schnoor, 2020). Yet aging is associated with declines in several cognitive abilities, including decision-making (Craik & Salthouse, 2011; Hess et al., 2015; Raz, 2000). In particular, older adults show impairments in complex decision-making requiring multi-step associations and judgments, such as forward planning (Drummond & Niv, 2020; Eppinger & Bruckner, 2015; Eppinger et al., 2013). However, the mechanism by which multi-step inference and planning decline with age is unclear, limiting efforts to develop effective interventions to reverse age-related cognitive decline. One challenge is that existing behavioral assays do not distinguish whether age-related declines arise from the judgment process itself or from difficulties learning the associations needed for such judgments, as in model-based planning (Feher da Silva & Hare, 2020; Konovalov & Krajbich, 2020).

Many real-world decisions require inferring structure from disparate experiences to form flexible, goal-directed plans. For instance, consider a student, Alice, preparing to present at an out-of-town conference. She books a 4 PM Friday flight and plans to call a taxi at noon to allow ample time for travel. Even if this is her first conference, she can engage in multi-step planning by assembling relevant knowledge from related experiences: after many delayed flights, she opts to arrive the night before her presentation (as opposed to the morning of); she anticipates heavier Friday afternoon traffic; and she chooses a taxi over driving due to past struggles finding airport parking.

Recent work in younger adults suggests that such goal-directed planning depends on inferring structure from the environment. Rmus et al., (2022) developed a task that measures how participants learn the associative structure among image pairs (edges) arranged along a latent “graph”. Participants implicitly learned complex associative structures (latent structure learning) from randomized exposures to individual graph edges. This knowledge was assessed via shortest path judgments requiring multi-step inference. Importantly, performance on the judgment task predicted greater use of model-based planning on a goal-directed decision-making task. A major strength of the task developed by Rmus et al. (2022) is its sensitivity to both the degree of latent structure learning and the ability to measure performance across varying associative distances. One open question, however, is how structured learning occurs, and what type of knowledge representations support multi-step inference judgments.

One reasonable assumption is that memory may influence how individuals learn latent structures from experience, as model-based planning involves hippocampal contributions and effective memory search (Bornstein & Daw, 2013; Doll et al., 2015; Vikbladh et al., 2019). In the task by Rmus et al., an accurate representation of individual associations is critical for learning the graph. More specifically, memory precision may be especially important for forming representations that enable efficient planning. For example, remembering Friday afternoon traffic patterns (rather than Monday afternoon or Friday evening) allows Alice to make more accurate predictions for her trip. In episodic memory, high precision can be achieved through pattern separation, in which competing information is encoded as distinct neural patterns (Bakker et al., 2008; Yassa & Stark, 2011). Pattern separation reduces or resolves memory interference, enabling discrimination of similar events (Lacy et al., 2011; Poppenk et al., 2013), but this ability declines with age (Burke et al., 2010; Stark et al., 2013; Toner et al., 2009; Yassa & Stark, 2011). Declines in pattern separation may explain older adults’ increased susceptibility to interference and related memory failures (Campbell et al., 2010; Wilson et al., 2006). We therefore asked whether age-related deficits in multi-step planning stem from impaired latent structure learning due to reduced precision and greater interference.

If older adults’ increased susceptibility to memory interference undermines their ability to learn the structures necessary for multistep planning, training interventions aimed at bolstering encoding may improve multi-step inference judgments needed for complex decision-making tasks. In support of this idea, work on episodic memory has shown that manipulating learning sequences or temporally separating overlapping associations can reduce memory interference and bias formation of distinct neural representations to support associative inference (Zeithamova & Preston, 2017; Zhou et al., 2023).

What is the nature of the representations that support structural inference?

Research suggests that associative inference can be supported by at least two kinds of neural representations (H. Eichenbaum, 2001; Howard Eichenbaum, 2017; Poppenk et al., 2013; Schlichting et al., 2015; Zhou et al., 2023). In standard associative inference tasks, participants learn A–B pairs, then B–C pairs, and are later tested on their knowledge of the indirect association between A and C, despite never having been presented with an A-C pair. If participants encode A-B and B-C as separate episodes during learning via pattern separation, they may form localist (orthogonalized) representations of each pair (Kumaran & McClelland, 2012; Zhou et al., 2023) and A–C inference requires effortful retrieval and recombination at test (e.g., “A paired with B, and B paired with C, so A is associated with C”). While localist representations yield high precision and resist interference, they may be less efficient for making inferences due to the additional effort required for information retrieval and recombination. (Zeithamova et al., 2012; Zhou et al., 2023).

Alternatively, encountering B–C may reactivate A–B, allowing integration of new (-C) information into an updated A–B–C representation (Morton et al., 2017; Schlichting & Preston, 2015; Zeithamova et al., 2012). These distributed representations support rapid inference but are prone to interference and false memories (Bowman et al., 2021; McCloskey & Cohen, 1989; Zhou et al., 2023), making them potentially challenging for individuals such as older adults who are already vulnerable to interference. If age-related declines in memory precision cause inference failures, then a learning method that reduces memory interference should improve latent structure learning and judgments.

Current Study

In the present study, we tested whether age-related declines in memory precision affect performance on a multi-step associative inference task (the “graph” task; (Rmus et al., 2022). Unlike standard associative inference tasks that use independent triads to test inference (e.g., AB-BC → AC; (Carpenter et al., 2021; Schlichting et al., 2015; Zhou et al., 2023), our task requires participants to integrate many overlapping associations into a coherent cognitive graph and to estimate relative distances within that graph. These representational operations have been shown to predict the degree to which individuals rely on model-based planning strategies (Rmus et al., 2022; Yoo et al., 2024). In Experiment 1, participants studied randomly presented object pairs drawn from an underlying graph of 12 nodes (objects) and 16 edges (object pairs). At test, they judged relative distances between object pairs (considered to be a fundamental computation for high-level planning; (Huang et al., 2019)) assessing their ability to mentally navigate associations and make shortest-path judgments. While the task does not involve overt goal-directed action or planning, the multi-step inference judgments directly probe the internal model-construction processes that support goal-directed planning. Accordingly, we interpret performance on our inference judgment test as indexing the fidelity of internal relational models that underlie overt planning behavior. We expected performance deficits with age, specifically driven by poor memory precision, so participants also completed the Mnemonic Discrimination Task (MDT) to independently measure mnemonic discrimination, a sensitive behavioral index of pattern separation ability that tracks age-related decline (Stark et al., 2019) and predicts decision-making beyond chronological age(Noh et al., 2023). In Experiment 2, we tested the prediction that separating overlapping edge pairs during learning (blocked training) would reduce working memory load and improve multi-step inference judgment performance for older adults or those with low memory precision (Schlichting et al., 2015). To confirm the relationship between training condition and memory precision, we simulated task performance using variants of artificial neural network models that differed only in their internal representational capacity and compared the model outputs to our behavioral findings.

Method

Experiments 1 and 2 were identical except for the learning-phase presentation order (intermixed vs. blocked). Data were collected concurrently, with participants randomly assigned to the intermixed condition (Experiment 1) or blocked condition (Experiment 2). Although the two training conditions were implemented under a single overarching protocol with random assignment of participants to conditions, we present the work as two experiments to reflect their distinct conceptual aims. Experiment 1’s task closely followed Rmus et al. (2022) but included a lifespan sample with older adults, was administered online, and incorporated an additional measure of memory specificity. Thus, Experiment 1 uses the original intermixed schedule (Rmus et al., 2022) to replicate prior findings and to test whether age or mnemonic discrimination better explains multi-step inference performance across the adult lifespan. Experiment 2 implements a temporally separated schedule to test the hypothesis that reducing interference would preferentially benefit low-precision learners. Anticipating that older adults might find the original task difficult, we implemented the blocked sequence in Experiment 2, based on prior work in associative memory suggesting that temporally separating overlapping associations might reduce cognitive load and improve performance in populations with lower memory capacity (Schlichting et al., 2015). Note that the “blocked” manipulation used here (i.e., separating overlapping edges across time over the course of training) is consistent with similar manipulations of “blocking” used in episodic memory (Beukers et al., 2024; Schlichting et al., 2015; Zhou et al., 2023) and is somewhat distinct from how “blocking” is used in the category learning literature (Brunmair & Richter, 2019; Kang & Pashler, 2012; Noh et al., 2024, 2016). Methods are described jointly below, with procedural differences between the two experiments noted explicitly.

Participants

Sample size was based on prior work with similar designs (N = 81; Rmus et al., 2022), with a ~25% increase to offset higher noise in our online data (relative to the supervised in-person data collected by Rmus et al., 2022). We aimed for ~100 participants per experiment post-exclusion. A total of 219 participants (112 female, 107 male; ages 19-84, mean[sd] age = 55.7[14.2]) were recruited online (see Supplementary Material; Online Data Collection): 113 in Experiment 1 (59 female, 54 male, ages 22-84, mean[sd] age = 56.7[13.8]), and 106 in Experiment 2 (53 female / 53 male, ages 19-79, mean[sd] age = 54.8[14.7]). Participants received monetary compensation and completed a tutorial plus a rotation-detection screener (10 practice trials; ≥70% accuracy required within two attempts to ensure attention to task instructions) to be eligible to participate in the full task. All protocols were classified as an “Exempt Online Survey” by the Institutional Review Board of the University of California, Irvine.

Procedure

Mnemonic Discrimination Task (MDT)

Participants completed the MST (Figure 1) as an independent measure of memory ability to assess whether mnemonic discrimination explained individual differences in graph task performance.

Figure 1. The Mnemonic Discrimination Task.

Figure 1.

Open in a new tab

(A) Participants view a sequence of objects during an incidental encoding phase in which participants are asked to classify each object as an indoor or outdoor object. (B) In a surprise discrimination test, participants view a series of objects and are asked to determine whether each object is an “old,” “new,” or “similar” item relative to what was shown during the encoding phase. (C) Relationship between chronological age and LDI across participants in both experiments. Shaded bands indicate 95% confidence intervals around the best fit regression line for each experiment (dashed lines).

Encoding Phase.

Participants viewed object images and made indoor/outdoor judgments (cover task). No MST images overlapped with those used in the graph task, and participants were unaware that their memory would later be tested.

Mnemonic Discrimination Test

On a surprise discrimination test, participants saw object images that were identical (old), similar (lures), or novel (foils) relative to objects shown during encoding, and judged each object as “old”, “similar”, or “new”. A Lure Discrimination Index (LDI) was calculated for each participant: p('sim'∣lure) - p('sim'∣foil). Higher LDI values indicate better ability to classify lures as “similar” relative to foils, reflecting greater mnemonic discrimination and better memory encoding precision (Stark et al., 2019).

Structural Inference (“Graph”) Task

Study Phase.

Participants viewed and were told to memorize a series of object pairs (Figure 2 Panel A), each presented for 1s in random order. Participants provided a rotation judgment on each trial to ensure attention. No information about the underlying structure was given, though pairs were drawn from a hidden 12-node, 16-edge graph (Figure 2 Panel B). In Experiment 1, all 16 unique edge pairs were repeated 44 times (704 trials total) in a random sequence (Figure 2 Panel C, “intermixed”). Experiment 2 grouped object pairs into 4 mini-blocks, each with 4 unique, non-overlapping pairs (Figure 2 Panel C, “blocked”). This separation of potentially confusing (overlapping) edges across time aimed to reduce memory interference and improve learning of the graph structure.

Figure 2. Structural Inference Task learning phase.

Figure 2.

Open in a new tab

Participants learn individual edges (A) drawn from a latent structure made up of several overlapping edges (B). In Experiment 1, edges are drawn randomly and presented in an intermixed fashion (C, intermixed), whereas in Experiment 2, overlapping edges are separated in time across different mini-blocks (C, blocked).

Following the study phase, participants completed two different tests to measure how well participants can use their knowledge of the graph structure (Figure 3 Panel A, Judgment Test), as well as how well they learned the graph structure (Figure 3 Panel B, Graph Reconstruction Test).

Figure 3. Structural Inference Task test phases.

Figure 3.

Open in a new tab

(A) After the learning phase, structural inference-based judgments were assessed for each participant. Participants were presented with 3 objects and asked whether the object on the left or right was closer to the center object, based on the associations they learned in the previous study phase. Judgment phase trials varied in difficulty based on the difference between choice options. The most difficult trials were ones in which options differed by an associative distance of 1 (distance 2 vs. 3, 3 vs. 4, or 4 vs. 5), whereas the easiest trials were ones in which choice options differed by an associative distance of 3 (distance 2 vs. 5). (B) After the judgment phase, participants were asked to reconstruct the graph to the best of their knowledge by placing all studied objects on a “canvas” on their screen and connecting objects only if they had been directly paired together during the study phase. Correctly drawn connections were classified as “hits,” whereas incorrectly drawn connections were classified as “false alarms.” Participants were required to place all objects on the canvas, and each object had to be connected to at least one other object to complete this phase.

Judgment Test.

After completing the learning phase, participants completed a relative distance judgment task. This task measured participants’ ability to use learned information for structural inference. Participants were asked to judge which of two objects (left vs. right) was closer to a central object, based on indirect relationships learned previously (i.e., graph distance). Responses were made using keyboard buttons, across 204 trials with up to 10 seconds per trial. Each unique object served as a central node 17 times. Choice options were randomly selected with the following constraints: 1) neither option was directly paired with the central node during the study, and 2) the shortest path length between the central node and the two options (associative distance) was not equal. Trial difficulty varied based on the difference in associative distance between choice options and the reference node (ranging from 1 to 3). Accuracy was calculated within each difficulty bin.

Because online data are unsupervised and noisier, judgment test data were screened for outliers prior to analysis using a two-step reaction time (RT)–based procedure, as RTs were not analyzed directly (see Supplementary Material; RT-based Filters for additional details). Filtered judgment-phase data were analyzed using linear mixed-effects models (R nlme package; (Pinheiro et al., 2017), with difficulty (distDiff = 1, 2, 3; 1= most difficult), age, LDI, and sequence (blocked vs. intermixed) specified as fixed effects, and participant as a random effect.

Graph Reconstruction Test.

Following the judgment test, participants completed a graph reconstruction task to assess explicit knowledge of the learned structure. All 12 study-phase objects were displayed above a blank canvas, and participants arranged them by clicking and dragging each object onto the canvas. Participants were asked to place and link objects that were directly paired during the study by selecting two objects sequentially; linked items were connected with a straight line. All objects had to be placed on the canvas and linked to at least one other object before submission.

We used a Signal Detection Theory approach to analyzing the reconstruction data (Carterette, 1967). We scored reconstructions as follows: “hits” for correctly linked study pairs, “false alarms” for incorrectly linked pairs, and the total number of edges drawn. Accuracy was computed in two ways: 1) sensitivity (d’), defined as “hit rate - false alarm rate”, and 2) precision. Across the 12 nodes, there are a total of 66 possible edges that could be drawn if every object pair were connected. Of these, 16 edges are true edges in the underlying graph structure (“hits”), and 50 represent non-edges (“false alarms”). Thus, the hit rate is calculated as hits/16, and the false-alarm rate is calculated as false alarms/50, and d′ is computed as hit rate minus false-alarm rate. We also examined an additional measure of accuracy (precision), taken from machine learning (Blair, 1979), to complement d′. While d′ separately considers hits and false alarms relative to the total possible true and false edges, precision instead evaluates the proportion of drawn edges that were correct (hits / total edges drawn). Precision is useful for comparing reconstruction accuracy across individuals who may differ in memory capacity. For example, an individual with high memory capacity may draw more edges overall, whereas a lower-capacity individual may draw fewer; precision helps interpret reconstruction accuracy in light of these differences. Thus, the precision metric adjusts performance based on the number of edges drawn, allowing for better comparisons across individuals with varying memory capacity.

Analyses were conducted in R using linear regression (lm). We separately modeled hits, false alarms, total edges, accuracy, and precision as a function of LDI, training sequence (blocked vs. intermixed), and their interaction. For each analysis, the full model was specified as follows: DV ~ LDI + Sequence + LDI*Sequence. If no significant interaction was found, the model was refitted without the interaction term to test for main effects (Grace-Martin, 2011).

Artificial Neural Network Model (ANN)

We implemented a set of simple five-layer feedforward autoencoders to simulate associative learning of the graph structure. The models learned to predict each object’s corresponding pair (as presented during study) and were then used to infer representational structure within their hidden layers. As with our human participants, half of the models were trained via the intermixed sequence, and the other half were trained via the blocked sequence. Across models, we also varied the number of units in the hidden layers to simulate differences in memory capacity (fewer units = lower memory capacity). This approach allowed us to test whether representational capacity could qualitatively reproduce the observed human interaction between memory precision and learning sequence. Full model specification, optimization parameters, and training details are provided in Supplementary Material.

Results

Experiment 1: Better memory encoding ability improves structural inference-based judgments

Judgment Test

We used a linear mixed-effects model to examine whether trial difficulty (relative distance difference between choice options) and chronological age predicted judgment accuracy (Figure 4): judgment accuracy ~ age + distDiff + age*distDiff, random = ~1∣subject (adjusted R2 = 0.23, AIC = −152.55). There was a main effect of age (βage = −0.030, SE = 0.012, t(110) = −2.57, p = 0.011), with older adults performing worse than younger adults. There was also a main effect of difficulty (βdistDiff = 0.032, SE = 0.009, t(216) = 3.50, p < 0.001), with performance improving as relative distance increased. The age × difficulty interaction was significant (βage*distDiff = −0.022, SE = 0.009, t(216) = −2.41, p = 0.017) such that younger adults showed greater performance gains as trials got easier (i.e., greater relative distance between options).

Figure 4. Behavioral results from Experiment 1.

Figure 4.

Open in a new tab

Judgment accuracy as a function of chronological age and trial difficulty (distance difference between choice options). The green band indicates 95% confidence intervals around the best fit regression line.

Adding mnemonic discrimination ability (LDI) to the model revealed it was a stronger predictor of judgment accuracy than age: judgment accuracy ~ age + distDiff + LDI + age*distDiff + age*LDI + distDiff*LDI + age*distDiff*LDI, random = ~1∣subject (adjusted R2 = 0.26, AIC = −133.98). When both age and LDI were included, the main effect of age (βage = −0.021, SE = 0.012, t(108) = −1.72, p = 0.088) and its interaction with difficulty (βage*distDiff = −0.012, SE = 0.010, t(214) = −1.24, p = 0.218) were no longer significant. Instead, higher LDI predicted better performance (βLDI = 0.031, SE = 0.013, t(108) = 2.42, p = 0.017), with a significant LDI × difficulty interaction (βLDI*distDiff = 0.033, SE = 0.010, t(214) = 3.34, p = 0.001) indicating that memory precision benefits emerged primarily in easier trials. Age × LDI and the three-way interaction were non-significant. Given LDI’s predictive strength, we re-ran the model excluding age (Figure 5, intermixed condition; adjusted R2 = 0.26, AIC = −164.19). This model confirmed significant main effects of LDI (βLDI = 0.035, SE = 0.012, t(110) = 2.99, p = 0.003) and difficulty (βdistDiff = 0.032, SE = 0.009, t(216) = 3.50, p < 0.001) as well as a robust LDI × difficulty interaction (βLDI*distDiff = 0.034, SE = 0.009, t(216) = 3.84, p < 0.001), with higher LDI predicting greater gains in easier trials.

Figure 5. Behavioral results combining data from Experiment 1 (blue) and Experiment 2 (red).

Figure 5.

Open in a new tab

Judgment accuracy as a function of trial difficulty (distance difference between choice options), mnemonic discrimination ability (LDI), and training condition (blocked vs. intermixed learning sequence). Shaded bands indicate 95% confidence intervals around the best fit regression line for each training condition (dashed lines).

These results suggest that structural inference performance depends on the precision of encoded associations, with the largest differences observed in trials requiring comparisons across larger associative distances (i.e., easier trials). Notably, participants with low memory abilities performed at chance across all difficulty levels, suggesting minimal or no latent structure learning. We hypothesized that this may reflect greater susceptibility to memory interference, which might be mitigated by training on non-overlapping subsets one block at a time. Thus, Experiment 2 introduced the “blocked” training intervention to reduce interference for low-LDI participants during learning and potentially improve subsequent judgment performance.

Experiment 2: Memory-based inference can be improved via individualized training

To test whether training conditions modulate structural inference as a function of memory ability, we combined data from Experiment 1 (n = 113, intermixed sequence) and Experiment 2 (n = 106, blocked sequence) and fit a linear mixed-effects model (Figure 5): judgment accuracy ~ LDI + distDiff + Sequence + LDI*distDiff + Sequence*LDI + Sequence*distDiff + LDI*Sequence*distDiff, random = ~1∣subject (adjusted R2 = 0.21, AIC = −304.55). There was a significant 3-way interaction between mnemonic discrimination ability, difficulty, and learning sequence (βLDI*distDiff*sequence = 0.054, SE = 0.013, t(422) = 4.10, p < 0.001), indicating that the relationship between mnemonic discrimination ability (LDI) and performance differed by training sequence and trial difficulty. Importantly, there was an interaction between learning sequence and mnemonic discrimination ability (βLDI*sequence = 0.038, SE = 0.016, t(214) = 2.26, p = 0.025): participants with low LDI benefitted from blocked training, whereas those with high LDI performed better with intermixed training. This crossover interaction was evident only in the easiest trials (distDiff = 3). Still, reducing memory interference via blocked training improved structural inference for individuals with weaker memory ability, enabling more accurate inference judgments when associative distances were large.

Artificial neural network captures relationship between training sequence and memory capacity in distance-dependent judgment performance

Given the observed interaction between training sequence and mnemonic discrimination ability (LDI) in Experiments 1 and 2, we tested whether such effects could be reproduced by manipulating a neural network’s representational capacity (Figure 6 Panel A). To test this, we implemented a five-layer feedforward autoencoder, which jointly optimizes encoding and decoding of sequential inputs (Blanco Malerba et al., 2024; Chandak et al., 2024), trained to associate paired items (edges) from the graph task (704 presentations; see Supplementary Material: Artificial Neural Network Model).

Figure 6. Neural network design and simulation results.

Figure 6.

Open in a new tab

(A) Schematic of two example artificial neural networks (the smallest and largest capacity models). To model memory capacities' effect on judgment accuracy, the width of the second and fourth layers was varied from 6 units (low-LDI; reflecting the hypothesized diminished capacity for separating inputs) to 36 units (high-LDI; allowing for sparse, highly separated representations of each input object). (B) Model judgment task results across varying layer widths and training conditions, matching the pattern observed in human participants with varying memory precision. Shaded bands indicate 95% confidence intervals around the best fit regression line for each training condition (dashed lines), and error bars around individual points reflect 95% confidence intervals around each mean.

To model variability in memory precision, we manipulated the number of units in the second and fourth layers (L2/L4; Figure 6 Panel A). Larger widths corresponded to more distinct, nonoverlapping representations of each node (analogous to higher LDI), whereas smaller widths simulated representational merging before latent structure extraction (analogous to lower LDI). Consistent with our behavioral results, there was a 3-way interaction between layer width, difficulty, and training sequence (βwidth*distDiff*sequence = 0.0032, SE = 0.0006, t(23996) = 5.58, p < 0.001) (Figure 6 Panel B). Additionally, there was an interaction between training sequence and the models’ layer width (βwidth*sequence = 0.0147, SE = 0.0008, t(11996) = 18.37, p < 0.001) such that structural inference-based judgments were optimized by different training conditions (blocked vs. intermixed learning sequences) depending on the neural network model’s capacity (larger vs. smaller hidden layer width) in a similar pattern to those seen in the human data (blocked > intermixed for low LDI, intermixed > blocked for high LDI). These interactions show that lower-capacity models benefited more from blocked training, while higher-capacity models benefited more from intermixed training–precisely the crossover pattern observed between memory precision and sequence in our human participants.

Graph Reconstruction.

To get a sense of the memory representations that participants may have formed during learning, we analyzed hits (correct edges), false alarms (incorrect edges), total edges drawn, and two accuracy metrics (sensitivity and precision) from the graph reconstruction phase as a function of mnemonic discrimination ability (LDI) and training sequence (blocked vs. intermixed). Each of these five reconstruction metrics (hits, false alarms, total edges, sensitivity, and precision) was used as the outcome measure with LDI and sequence as independent predictors (outcome ~ LDI + Sequence + LDI*Sequence). Across all five regressions, the LDI × Sequence interaction was not significant, so interaction terms were removed and models were refit to assess main effects; the results reported here are from those main-effects models. For hits, the linear regression (hits ~ LDI + Sequence) was significant, adjusted R2 = .03, F(2, 212) = 4.09, p = .018, with a main effect of LDI, β = 0.65, SE = 0.23, t(212) = 2.81, p = .005, indicating that participants with better memory ability produced more correct edges; there was no effect of training sequence (Figure 7 Panel A). For false alarms, the model (false alarms ~ LDI + Sequence) was significant, adjusted R2 = .03, F(2, 212) = 4.07, p = .018, with a main effect of training sequence, β = 1.53, SE = 0.55, t(212) = 2.77, p = .006, such that participants in the intermixed condition produced more false edges; LDI was not significant (Figure 7 Panel B). For total edges, the model (total edges ~ LDI + Sequence) was significant, adjusted R2 = .05, F(2, 212) = 6.38, p = .002, with main effects of LDI, β = 0.79, SE = 0.33, t(212) = 2.36, p = .019, and Sequence, β = 1.71, SE = 0.67, t(212) = 2.55, p = .011, indicating that higher-LDI participants and those in the intermixed condition drew more edges overall (Figure 7 Panel C). We also assessed reconstruction accuracy using two metrics: sensitivity (proportion of hits - proportion of false alarms) and precision (number of hits divided by the total number of drawn edges). For sensitivity, the model (sensitivity ~ LDI + Sequence) was significant, adjusted R2 = .02, F(2, 212) = 3.12, p = .046, revealing a main effect of LDI, β = 0.04, SE = 0.01, t(212) = 2.49, p = .014, with higher-LDI participants showing better reconstruction accuracy (Figure 7 Panel D). For precision, which adjusts for the tendency of low-memory participants to draw fewer edges, the model (precision ~ LDI + Sequence) was marginal, adjusted R2 = .02, F(2, 212) = 3.03, p = .050, with trending effects of LDI, β = 0.04, SE = 0.02, t(212) = 1.72, p = .088, and Sequence, β = –0.09, SE = 0.05, t(212) = –1.86, p = .065, suggesting higher precision for participants with better memory ability and for those trained in the blocked condition (Figure 7 Panel E).

Figure 7. Behavioral results combining data from Experiment 1 (blue) and Experiment 2 (red) for the graph reconstruction phase.

Figure 7.

Open in a new tab

Shaded bands indicate 95% confidence intervals around the best fit regression line for each training condition (dashed lines).

Discussion

The present study examined whether multi-step inference judgments requiring latent structure learning are affected by age-related cognitive decline. In Experiment 1, we found that older adults performed worse on structural inference judgments, but these deficits were better explained by individual differences in mnemonic discrimination ability (LDI) than by chronological age. Experiment 2 demonstrated that structural inference can be improved through individualized training: intermixing overlapping associations benefited those with high LDI, while blocked training benefited those with low LDI. These behavioral findings were reinforced by artificial neural network model simulations that varied in memory capacity to examine multi-step inference performance under blocked and interleaved training conditions. Although the models are a simplified system, manipulating representational capacity in these networks successfully reproduced the sequence x memory precision interaction observed in human participants, thereby providing converging evidence for our hypotheses.

Collectively, our findings suggest that the way individuals encode and organize information interacts with memory ability to shape representations that vary in their effectiveness for supporting accurate inference. Matching training to an individual’s representational tendencies may be critical for optimizing judgment performance. Although our task did not require participants to execute planned sequences of actions, the multi-step inference judgments it elicited engage the representational mechanisms thought to support model-based planning (Rmus et al., 2022; Yoo et al., 2024). Specifically, the judgment test requires estimating which of two stimuli is “closer” within a learned associative graph, effectively querying the internal relational model used to chain together multiple steps of the structure. Prior work using similar graph-learning tasks has demonstrated that the integrity of these relational models predicts individuals’ reliance on model-based strategies (Rmus et al., 2022). Thus, our findings offer insight into how age- and memory-related variability in representational fidelity may constrain the computations that underlie goal-directed planning.

In Experiment 1, evidence of latent structure learning emerged only in participants with high mnemonic discrimination ability. High-LDI individuals achieved above-chance inference performance, especially on easier trials, whereas low-LDI participants performed at chance. The lack of learning observed in low-LDI participants who performed our task may explain why certain populations such as older adults and individuals with mild cognitive impairments may exhibit deficits in goal-directed planning (Eppinger & Bruckner, 2015; Samanez-Larkin, 2013), as prior work shows that these populations are particularly vulnerable to deficits in memory precision (Yassa & Stark, 2011; Yassa et al., 2010). High-LDI participants also produced more accurate graph reconstructions across multiple metrics (number of correct edges, reconstruction accuracy measured using sensitivity and precision), but they also generated more false alarms and total edges. One explanation is that making inference judgments before reconstruction induced false memories by misattributing inferred information as directly experienced (Bowman et al., 2021). However, this order effect should have applied equally to all participants, making it unlikely to explain the selective increase in false alarms for high-LDI individuals. A more plausible account is that high-LDI individuals tend to form distributed representations during learning (Kumaran & McClelland, 2012; Zhou et al., 2023). According to parallel distributed processing models such as C-HORSE, distributed representations support flexible inference but are prone to false alarms (Zhou et al., 2023). Our data fit this prediction: high-LDI participants outperformed others on inference judgments but also produced more false edges.

In Experiment 2, we tested whether separating overlapping pairs in time (blocked training) would reduce memory interference and improve inference for low-LDI individuals. Indeed, blocked training improved structural inference for low-LDI participants but impaired performance for high-LDI participants, producing a crossover interaction. Visual inspection of graph reconstructions of the best performers in each condition (Supplementary Figure 1) suggests why: high-LDI participants in the intermixed condition tended to create highly interconnected, distributed graphs, whereas blocked training led them to form more fragmented, localist graphs. In contrast, low-LDI participants generally produced localized graphs, but blocked training encouraged more interconnections relative to intermixing. These patterns suggest that memory encoding ability may bias representational style—distributed for high-LDI, localist for low-LDI—and that training sequences aligning with these preferences yield better performance. This hints at the idea that blocked vs. interleaved learning sequences can be leveraged to improve multi-step inference and other forms of inferential reasoning as a function of one’s memory capacity.

This framework may help reconcile conflicting findings in associative memory research. Some studies report that intermixing promotes distributed representations (Zhou et al., 2023), whereas others find it encourages more localist coding (Schlichting et al., 2015) with the reverse pattern for blocked learning. A key methodological difference is that Zhou et al.’s “blocked” condition still included intermixed trials within the same learning phase, likely negating interference-reduction benefits. This design may have disproportionately favored high-LDI participants while preventing low-LDI learners from reaching performance levels sufficient to influence overall effects. Supporting this, Zhou et al. excluded many low performers, yet performance remained low relative to Schlichting et al.

Recent neural network modeling shows that distributed representations are more likely to emerge when information is presented in an intermixed sequence (Zhou et al., 2023). Models with distributed coding predict that blocked learning increases memory interference as new information is acquired (McCloskey & Cohen, 1989; Zhou et al., 2023). Our behavioral data support this: the high-LDI intermixed group showed the best inference performance but also the highest false alarms during reconstruction—consistent with distributed coding—while the high-LDI blocked group exhibited low false alarms (Figure 7 Panel B) and high graph reconstruction precision (Figure 7 Panel E), suggesting more localist coding. However, this gain in precision came at the cost of impaired judgment performance (Figure 5). This tradeoff parallels evidence that temporally contiguous presentation facilitates integrative encoding (Pudhiyidath et al., 2022; Schapiro et al., 2013; Zeithamova & Preston, 2017; Zhou et al., 2023), whereas separating overlapping information can disrupt inference if it requires linking across longer temporal gaps (Zeithamova & Preston, 2017; Zhou et al., 2023). Visual inspection supports this: high-LDI blocked learners’ graphs were less cohesive than those in the high-LDI intermixed group (Supplementary Figure 1), consistent with blocking fragmenting an otherwise distributed network into more localized subgraphs, making cross-representation inference harder.

For low-LDI learners—such as older adults with age-related decline—susceptibility to memory interference can severely impair learning and memory encoding in an intermixed schedule. With 16 overlapping pairs, interference may be so severe that learners can only focus on memorizing a small subset of pairs—similar to older adults’ strategy of selectively encoding high-value items (Castel, 2007; Castel et al., 2002). Indeed, the top performers in the low-LDI intermixed condition reconstructed only the minimum six edges required to complete the task, suggesting encoding of a few localized subsets. Since our graph task likely benefits from distributed representations, the main challenge for low-LDI learners is overcoming interference during encoding (Kirwan & Stark, 2007; Pettigrew & Martin, 2014; Shimamura & Jurica, 1994). Blocked training may mitigate this by spacing overlapping pairs across time, allowing learners to consolidate subsets (e.g., A–B) before encountering related pairs (e.g., B–C). Previous studies in episodic memory have also suggested that stabilizing A-B representations prior to introducing overlapping B-C episodes can prevent memory interference and promote integration of the related episodes (Noh et al., 2026; Schlichting et al., 2015). Once initial representations are stabilized through repetition, new overlapping pairs can be integrated into existing knowledge (A–B–C) by pattern completing the overlapping memory traces (Schlichting et al., 2015). Consistent with this account, low-LDI participants in the blocked condition showed better inference judgments (Figure 5) and fewer false alarms (Figure 7 Panel B), indicating reduced interference. This pattern was also replicated with the ANN model simulations, showing that smaller-capacity models performed better under blocked (vs. intermixed) training conditions. Visual reconstructions suggest blocked training promoted more integrated networks in low-LDI learners than intermixing, consistent with prior work showing that blocking overlapping pairs fosters integrated representations, while intermixing promotes pattern-separated representations (Schlichting et al., 2015).

Collectively, these results indicate that structural inference depends on both memory encoding ability and the alignment of training structure with representational tendencies. For high-LDI learners, intermixing supports distributed networks that facilitate inference; for low-LDI learners, blocking reduces interference and fosters integration. These findings suggest that some age-related deficits in model-based planning and decision-making may stem from failures to form adequate latent structures. Because multi-step inference over a cognitive graph provides a substrate for evaluating relational distances and future action paths, these findings have direct implications for models of goal-directed decision making. The present task therefore contributes to understanding how aging- and memory-related representational constraints shape the cognitive computations underlying planning, even in the absence of overt behavioral choice. Future work should test how disruptions in latent structure learning impair multi-step planning (Harhen & Bornstein, 2023; Yoo et al., 2024), directly measure memory for individual associations, and use neuroimaging to track the formation of representations during learning. Given that successful inference can emerge from different representational formats, future studies should also examine their flexibility and limitations across different inference and planning contexts. Despite limitations, our results offer a mechanistic basis for individualized learning interventions to mitigate cognitive decline effects on decision-making across the lifespan.

Supplementary Material

supp

Funding

This work was supported by the National Institute of Aging at the National Institutes of Health (R21AG072673 and R01AG088306 to AMB, R01AG066683 and P30AG066519 to CELS, and F32AG072836 to SMN); and the National Science Foundation (NSF DGE-1839285 to KWC).

Footnotes

Conflict of Interest

The authors declare no competing financial interests.

Data Availability

All de-identified data and the final versions of all R scripts used to generate each data figure, as well as Pytorch model code, are freely available at the UCI CCNL GitHub repository (https://github.com/uciccnl). This study was not preregistered.

References

  1. Akhtar A. (2019, November 1). Corporate America is seeing a spike in the age of CEOs being hired - and yes, they’re overwhelmingly white men. Business Insider. https://www.businessinsider.in/strategy/news/corporate-america-is-seeing-a-spike-in-the-age-of-ceos-being-hired-and-yes-theyre-overwhelmingly-white-men/articleshow/71856494.cms [Google Scholar]
  2. Bakker A, Kirwan CB, Miller M, & Stark CEL (2008). Pattern separation in the human hippocampal CA3 and dentate gyrus. Science, 319(5870), 1640–1642. 10.1126/science.1152882 [DOI] [PMC free article] [PubMed] [Google Scholar]
  3. Beukers AO, Collin SHP, Kempner RP, Franklin NT, Gershman SJ, & Norman KA (2024). Blocked training facilitates learning of multiple schemas. Communications Psychology, 2(1), 28. 10.1038/s44271-024-00079-4 [DOI] [PMC free article] [PubMed] [Google Scholar]
  4. Blair DC (1979). Information Retrieval, 2nd ed. C.J. Van Rijsbergen. London: Butterworths; 1979: 208 pp. Price: $32.50. Journal of the American Society for Information Science. American Society for Information Science, 30(6), 374–375. 10.1002/asi.4630300621 [DOI] [Google Scholar]
  5. Blanco Malerba S, Micheli A, Woodford M, & Azeredo da Silveira R. (2024). Jointly efficient encoding and decoding in neural populations. PLoS Computational Biology, 20(7), e1012240. 10.1371/journal.pcbi.1012240 [DOI] [PMC free article] [PubMed] [Google Scholar]
  6. Bornstein AM, & Daw ND (2013). Cortical and hippocampal correlates of deliberation during model-based decisions for rewards in humans. PLoS Computational Biology, 9(12), e1003387. 10.1371/journal.pcbi.1003387 [DOI] [PMC free article] [PubMed] [Google Scholar]
  7. Bowman CR, de Araujo Sanchez M-A, Hou W, Rubin S, & Zeithamova D. (2021). Generalization and False Memory in an Acquired Equivalence Paradigm: The Influence of Physical Resemblance Across Related Episodes. Frontiers in Psychology, 12, 669481. 10.3389/fpsyg.2021.669481 [DOI] [PMC free article] [PubMed] [Google Scholar]
  8. Brunmair M, & Richter T. (2019). Similarity matters: A meta-analysis of interleaved learning and its moderators. Psychological Bulletin, 145(11), 1029–1052. 10.1037/bul0000209 [DOI] [PubMed] [Google Scholar]
  9. Burke SN, Wallace JL, Nematollahi S, Uprety AR, & Barnes CA (2010). Pattern separation deficits may contribute to age-associated recognition impairments. Behavioral Neuroscience, 124(5), 559–573. 10.1037/a0020893 [DOI] [PMC free article] [PubMed] [Google Scholar]
  10. Campbell KL, Hasher L, & Thomas RC (2010). Hyper-binding: a unique age effect. Psychological Science, 21(3), 399–405. 10.1177/0956797609359910 [DOI] [PMC free article] [PubMed] [Google Scholar]
  11. Carpenter AC, Thakral PP, Preston AR, & Schacter DL (2021). Reinstatement of item-specific contextual details during retrieval supports recombination-related false memories. NeuroImage, 236(118033), 118033. 10.1016/j.neuroimage.2021.118033 [DOI] [PMC free article] [PubMed] [Google Scholar]
  12. Carterette EC (1967). Toward a more rigorous psychophysics: Signal detection theory and psychophysics . David M. green and John A. swets. Wiley, New York, 1966. 467 pp., illus. $12.95. Science, 156(3775), 632–633. 10.1126/science.156.3775.632 [DOI] [Google Scholar]
  13. Castel AD (2007). The Adaptive and Strategic Use of Memory By Older Adults: Evaluative Processing and Value-Directed Remembering. In Benjamin AS & Ross BH (Eds.), Psychology of Learning and Motivation (Vol. 48, pp. 225–270). Academic Press. 10.1016/S0079-7421(07)48006-9 [DOI] [Google Scholar]
  14. Castel AD, Benjamin AS, Craik FIM, & Watkins MJ (2002). The effects of aging on selectivity and control in short-term recall. Memory & Cognition, 30(7), 1078–1085. 10.3758/bf03194325 [DOI] [PubMed] [Google Scholar]
  15. Chandak S, Shah P, Borkar VS, & Dodhia P. (2024). Reinforcement learning in non-Markovian environments. Systems & Control Letters, 185(105751), 105751. 10.1016/j.sysconle.2024.105751 [DOI] [Google Scholar]
  16. Craik FIM, & Salthouse TA (2011). The Handbook of Aging and Cognition: Third Edition. Psychology Press. https://play.google.com/store/books/details?id=YeJ4AgAAQBAJ [Google Scholar]
  17. Doll BB, Shohamy D, & Daw ND (2015). Multiple memory systems as substrates for multiple decision systems. Neurobiology of Learning and Memory, 117, 4–13. 10.1016/j.nlm.2014.04.014 [DOI] [PMC free article] [PubMed] [Google Scholar]
  18. Drummond N, & Niv Y. (2020). Model-based decision making and model-free learning. Current Biology: CB, 30(15), R860–R865. 10.1016/j.cub.2020.06.051 [DOI] [PubMed] [Google Scholar]
  19. Eichenbaum H. (2001). The hippocampus and declarative memory: cognitive mechanisms and neural codes. Behavioural Brain Research, 127(1–2), 199–207. 10.1016/s0166-4328(01)00365-5 [DOI] [PubMed] [Google Scholar]
  20. Eichenbaum Howard. (2017). Memory: Organization and Control. Annual Review of Psychology, 68, 19–45. 10.1146/annurev-psych-010416-044131 [DOI] [Google Scholar]
  21. Eppinger B, & Bruckner R. (2015). Chapter 4 - Towards a Mechanistic Understanding of Age-Related Changes in Learning and Decision Making: A Neuro-Computational Approach. In Hess TM, Strough J, & Löckenhoff CE (Eds.), Aging and Decision Making (pp. 61–77). Academic Press. 10.1016/B978-0-12-417148-0.00004-2 [DOI] [Google Scholar]
  22. Eppinger B, Walter M, Heekeren HR, & Li S-C (2013). Of goals and habits: age-related and individual differences in goal-directed decision-making. Frontiers in Neuroscience, 7, 253. 10.3389/fnins.2013.00253 [DOI] [PMC free article] [PubMed] [Google Scholar]
  23. Feher da Silva C, & Hare TA (2020). Humans primarily use model-based inference in the two-stage task. Nature Human Behaviour, 4(10), 1053–1066. 10.1038/s41562-020-0905-y [DOI] [Google Scholar]
  24. Grace-Martin K. (2011, April 2). Testing and dropping interaction terms in regression and ANOVA models. The Analysis Factor. https://www.theanalysisfactor.com/testing-and-dropping-interaction-terms/ [Google Scholar]
  25. Hall M, & Hickey W. (2022, September 14). This one chart shows an unprecedented recent spike in age for Congress, posing a huge problem for American democracy. Business Insider. https://www.businessinsider.com/congress-oldest-history-gerontocracy-lawmakers-2022-9 [Google Scholar]
  26. Harhen NC, & Bornstein AM (2023). Overharvesting in human patch foraging reflects rational structure learning and adaptive planning. Proceedings of the National Academy of Sciences of the United States of America, 120(13), e2216524120. 10.1073/pnas.2216524120 [DOI] [PMC free article] [PubMed] [Google Scholar]
  27. Hess TM, Strough J, & Löckenhoff C. (2015). Aging and Decision Making: Empirical and Applied Perspectives. Academic Press. [Google Scholar]
  28. Huang Z, Liu F, & Su H. (2019). Mapping state space using landmarks for universal goal reaching. arXiv. 10.48550/arXiv.1908.05451 [DOI] [Google Scholar]
  29. Ingraham C. (2014, April 1). Yes, Congress is getting older – but so are the rest of us. The Washington Post. https://www.washingtonpost.com/news/wonk/wp/2014/04/01/yes-congress-is-getting-older-but-so-are-the-rest-of-us/ [Google Scholar]
  30. Kang SHK, & Pashler H. (2012). Learning Painting Styles: Spacing is Advantageous when it Promotes Discriminative Contrast. Applied Cognitive Psychology, 26(1), 97–103. 10.1002/acp.1801 [DOI] [Google Scholar]
  31. Kingma DP, & Ba J. (2014). Adam: A Method for Stochastic Optimization. arXiv. 10.48550/arXiv.1412.6980 [DOI] [Google Scholar]
  32. Kirwan CB, & Stark CEL (2007). Overcoming interference: an fMRI investigation of pattern separation in the medial temporal lobe. Learning & Memory, 14(9), 625–633. 10.1101/lm.663507 [DOI] [PMC free article] [PubMed] [Google Scholar]
  33. Konovalov A, & Krajbich I. (2020). Mouse tracking reveals structure knowledge in the absence of model-based choice. Nature Communications, 11(1), 1893. 10.1038/s41467-020-15696-w [DOI] [Google Scholar]
  34. Kumaran D, & McClelland JL (2012). Generalization through the recurrent interaction of episodic memories: a model of the hippocampal system. Psychological Review, 119(3), 573–616. 10.1037/a0028681 [DOI] [PMC free article] [PubMed] [Google Scholar]
  35. Lacy JW, Yassa MA, Stark SM, Muftuler LT, & Stark CEL (2011). Distinct pattern separation related transfer functions in human CA3/dentate and CA1 revealed using high-resolution fMRI and variable mnemonic similarity. Learning & Memory, 18(1), 15–18. 10.1101/lm.1971111 [DOI] [PMC free article] [PubMed] [Google Scholar]
  36. McCloskey M, & Cohen NJ (1989). Catastrophic Interference in Connectionist Networks: The Sequential Learning Problem. In Bower GH (Ed.), Psychology of Learning and Motivation (Vol. 24, pp. 109–165). Academic Press. 10.1016/S0079-7421(08)60536-8 [DOI] [Google Scholar]
  37. Morton NW, Sherrill KR, & Preston AR (2017). Memory integration constructs maps of space, time, and concepts. Current Opinion in Behavioral Sciences, 17, 161–168. 10.1016/j.cobeha.2017.08.007 [DOI] [PMC free article] [PubMed] [Google Scholar]
  38. Noh SM, Bjork RA, & Preston AR (2024). General knowledge and detailed memory benefit from different training sequences. Journal of Applied Research in Memory and Cognition, 13(3), 329–341. 10.1037/mac0000193 [DOI] [PMC free article] [PubMed] [Google Scholar]
  39. Noh SM, Singla UK, Bennett IJ, & Bornstein AM (2023). Memory precision and age differentially predict the use of decision-making strategies across the lifespan. Scientific Reports, 13(1), 17014. 10.1038/s41598-023-44107-5 [DOI] [PMC free article] [PubMed] [Google Scholar]
  40. Noh SM, Yan VX, Bjork RA, & Maddox WT (2016). Optimal sequencing during category learning: Testing a dual-learning systems perspective. Cognition, 155, 23–29. 10.1016/j.cognition.2016.06.007 [DOI] [PMC free article] [PubMed] [Google Scholar]
  41. Noh SM, Zhou D, Cooper KW, Guo S, Dinh ET, & Bornstein AM (2026). Sparsity and memory constraints interact with training sequence to bias learning of associative maps. Neuropsychologia, 109396, 109396. 10.1016/j.neuropsychologia.2026.109396 [DOI] [Google Scholar]
  42. Ogletree AM, & Katz B. (2021). How Do Older Adults Recruited Using MTurk Differ From Those in a National Probability Sample? International Journal of Aging & Human Development, 93(2), 700–721. 10.1177/0091415020940197 [DOI] [PubMed] [Google Scholar]
  43. Paszke A, Gross S, Massa F, Lerer A, Bradbury J, Chanan G, Killeen T, Lin Z, Gimelshein N, Antiga L, Desmaison A, Köpf A, Yang E, DeVito Z, Raison M, Tejani A, Chilamkurthy S, Steiner B, Fang L, … Chintala S. (2019). PyTorch: An imperative style, high-performance deep learning library. arXiv. 10.48550/arXiv.1912.01703 [DOI] [Google Scholar]
  44. Pettigrew C, & Martin RC (2014). Cognitive declines in healthy aging: evidence from multiple aspects of interference resolution. Psychology and Aging, 29(2), 187–204. 10.1037/a0036085 [DOI] [PubMed] [Google Scholar]
  45. Pinheiro J, Bates D, DebRoy S, Sarkar D, Heisterkamp S, Van Willigen B, & Maintainer R. (2017). Package ‘nlme’. Linear and Nonlinear Mixed Effects Models, Version, 3(1). http://cran.rapporter.net/web/packages/nlme/nlme.pdf [Google Scholar]
  46. Poppenk J, Evensmoen HR, Moscovitch M, & Nadel L. (2013). Long-axis specialization of the human hippocampus. Trends in Cognitive Sciences, 17(5), 230–240. 10.1016/j.tics.2013.03.005 [DOI] [PubMed] [Google Scholar]
  47. Pudhiyidath A, Morton NW, Duran RV, Schapiro AC, Momennejad I, Hinojosa-Rowland DM, Molitor RJ, & Preston AR (2022). Representations of temporal community structure in hippocampus and precuneus predict inductive reasoning decisions. Journal of Cognitive Neuroscience, 34(10), 1736–1760. 10.1162/jocn_a_01864 [DOI] [PMC free article] [PubMed] [Google Scholar]
  48. Ratcliff R, & Hendrickson AT (2021). Do data from mechanical Turk subjects replicate accuracy, response time, and diffusion modeling results? Behavior Research Methods, 53(6), 2302–2325. 10.3758/s13428-021-01573-x [DOI] [PMC free article] [PubMed] [Google Scholar]
  49. Raz N. (2000). Aging of the brain and its impact on cognitive performance: Integration of structural and functional findings. In Craik FIM & Salthouse TA (Eds.), The handbook of aging and cognition (2nd ed., pp. 1–90). Lawrence Erlbaum Associates Publishers. [Google Scholar]
  50. Rmus M, Ritz H, Hunter LE, Bornstein AM, & Shenhav A. (2022). Humans can navigate complex graph structures acquired during latent learning. Cognition, 225, 105103. 10.1016/j.cognition.2022.105103 [DOI] [PMC free article] [PubMed] [Google Scholar]
  51. Samanez-Larkin GR (2013). Financial Decision Making and the Aging Brain. APS Observer, 26(5), 30–33. https://www.ncbi.nlm.nih.gov/pubmed/23946614 [PMC free article] [PubMed] [Google Scholar]
  52. Schapiro AC, Rogers TT, Cordova NI, Turk-Browne NB, & Botvinick MM (2013). Neural representations of events arise from temporal community structure. Nature Neuroscience, 16(4), 486–492. 10.1038/nn.3331 [DOI] [PMC free article] [PubMed] [Google Scholar]
  53. Schlichting ML, Mumford JA, & Preston AR (2015). Learning-related representational changes reveal dissociable integration and separation signatures in the hippocampus and prefrontal cortex. Nature Communications, 6, 8151. 10.1038/ncomms9151 [DOI] [Google Scholar]
  54. Schlichting ML, & Preston AR (2015). Memory integration: neural mechanisms and implications for behavior. Current Opinion in Behavioral Sciences, 1, 1–8. 10.1016/j.cobeha.2014.07.005 [DOI] [PMC free article] [PubMed] [Google Scholar]
  55. Schnoor A. (2020, February 6). The Problem of the Aging CEO. Medium: The Intelligence of Everything. https://medium.com/the-intelligence-of-everything/the-problem-of-the-aging-ceo-e09e8d21bb65
  56. Shimamura AP, & Jurica PJ (1994). Memory interference effects and aging: Findings from a test of frontal lobe function. Neuropsychology, 8(3), 408–412. 10.1037/0894-4105.8.3.408 [DOI] [Google Scholar]
  57. Stark SM, Kirwan CB, & Stark CEL (2019). Mnemonic Similarity Task: A Tool for Assessing Hippocampal Integrity. Trends in Cognitive Sciences, 23(11), 938–951. 10.1016/j.tics.2019.08.003 [DOI] [PMC free article] [PubMed] [Google Scholar]
  58. Stark SM, Yassa MA, Lacy JW, & Stark CEL (2013). A task to assess behavioral pattern separation (BPS) in humans: Data from healthy aging and mild cognitive impairment. Neuropsychologia, 51(12), 2442–2449. 10.1016/j.neuropsychologia.2012.12.014 [DOI] [PMC free article] [PubMed] [Google Scholar]
  59. Toner CK, Pirogovsky E, Kirwan CB, & Gilbert PE (2009). Visual object pattern separation deficits in nondemented older adults. Learning & Memory, 16(5), 338–342. 10.1101/lm.1315109 [DOI] [PubMed] [Google Scholar]
  60. Turner AM, Engelsma T, Taylor JO, Sharma RK, & Demiris G. (2020). Recruiting older adult participants through crowdsourcing platforms: Mechanical Turk versus Prolific Academic. AMIA Annual Symposium Proceedings, 2020, 1230–1238. https://www.ncbi.nlm.nih.gov/pubmed/33936499 [Google Scholar]
  61. Vikbladh OM, Meager MR, King J, Blackmon K, Devinsky O, Shohamy D, Burgess N, & Daw ND (2019). Hippocampal Contributions to Model-Based Planning and Spatial Memory. Neuron, 102(3), 683–693.e4. 10.1016/j.neuron.2019.02.014 [DOI] [PMC free article] [PubMed] [Google Scholar]
  62. Wilson IA, Gallagher M, Eichenbaum H, & Tanila H. (2006). Neurocognitive aging: prior memories hinder new hippocampal encoding. Trends in Neurosciences, 29(12), 662–670. 10.1016/j.tins.2006.10.002 [DOI] [PMC free article] [PubMed] [Google Scholar]
  63. Yassa MA, & Stark CEL (2011). Pattern separation in the hippocampus. Trends in Neurosciences, 34(10), 515–525. 10.1016/j.tins.2011.06.006 [DOI] [PMC free article] [PubMed] [Google Scholar]
  64. Yassa MA, Stark SM, Bakker A, Albert MS, Gallagher M, & Stark CEL (2010). High-resolution structural and functional MRI of hippocampal CA3 and dentate gyrus in patients with amnestic Mild Cognitive Impairment. NeuroImage, 51(3), 1242–1252. 10.1016/j.neuroimage.2010.03.040 [DOI] [PMC free article] [PubMed] [Google Scholar]
  65. Yoo J, Chrastil ER, & Bornstein AM (2024). Cognitive graphs: Representational substrates for planning. Decision, 11(4), 537–556. 10.1037/dec0000249 [DOI] [PMC free article] [PubMed] [Google Scholar]
  66. Zeithamova D, & Preston AR (2017). Temporal Proximity Promotes Integration of Overlapping Events. Journal of Cognitive Neuroscience, 29(8), 1311–1323. 10.1162/jocn_a_01116 [DOI] [PMC free article] [PubMed] [Google Scholar]
  67. Zeithamova D, Schlichting ML, & Preston AR (2012). The hippocampus and inferential reasoning: building memories to navigate future decisions. Frontiers in Human Neuroscience, 6, 70. 10.3389/fnhum.2012.00070 [DOI] [PMC free article] [PubMed] [Google Scholar]
  68. Zhou Z, Singh D, Tandoc MC, & Schapiro AC (2023). Building integrated representations through interleaved learning. Journal of Experimental Psychology. General, 152(9), 2666–2684. 10.1037/xge0001415 [DOI] [PubMed] [Google Scholar]

Associated Data

This section collects any data citations, data availability statements, or supplementary materials included in this article.

Supplementary Materials

supp
NIHMS2161888-supplement-supp.docx (1.8MB, docx)

Data Availability Statement

All de-identified data and the final versions of all R scripts used to generate each data figure, as well as Pytorch model code, are freely available at the UCI CCNL GitHub repository (https://github.com/uciccnl). This study was not preregistered.

RESOURCES