An Exploration of Individual and Collective Reversal Learning in Rats

Abstract

Although associative learning research has been conducted for more than a century, little is known about learning processes when subjects are not alone, but in a group—a phenomenon termed collective learning. In collective learning situations, the behavior of conspecifics may serve as an associative cue for learning, like any other stimulus during individual learning. Two experiments investigated how individual versus collective training affects associative learning. Experiment 1 utilized a simultaneous discrimination task, whereas Experiment 2 implemented a serial go/no-go discrimination task. In both experiments, rats were trained either individually or collectively, exposing them to two distinct stimuli with only one of them signaling the availability of food reinforcement. Following acquisition training, all rats were tested both individually and collectively. Contingencies were then reversed: the previously nonreinforced stimulus now signaled the availability of food, and the previously reinforced stimulus now signaled the absence of food. Following reversal training, the rats were again tested individually and collectively. Results from both experiments suggest that the training condition (individual or collective) had little effect on learning the cue–outcome association. However, individual training negatively affected test performance in a collective context. These results suggest that collective training may have a facilitative effect on learning and points out key methodological considerations for more in-depth examination of this effect.

Supplementary Information

The online version contains supplementary material available at 10.1007/s40614-025-00450-8.

De Houwer and Hughes (2023) define learning as “changes in the way a system behaves toward a stimulus that result from regularities in the environment of that system” (p. 650). They use the term system in this definition to expand the scope of learning agents to include, among others, groups of animals. Nonetheless, associative learning models are almost invariantly focused on how individual organisms learn in isolation (e.g. Collins & Shenhav, 2022), neglecting the fact that many social animals live and may even learn in groups.

Collective learning is a process in which the discoveries and actions of individual organisms change the way in which the other members of a group behave (Laland & Plotkin, 1990; Nagy et al., 2020; Sumpter, 2006). This form of learning is observed in many types of organisms such as ants sharing information to enhance group foraging efficiency (Sasaki & Pratt, 2012), to humans collaborating to complete memory tests (Laughlin, 2011). Previous work has shown that the benefits of collective learning may depend on the task, whether it involves food foraging or associative learning tasks, and on the size of the group (Beauchamp & Kacelnik, 1991; Laland & Plotkin, 1990; Sumpter, 2006). However, much of this work has focused on when and how focal individual animals obtain information from other group members (Camacho-Alpízar & Guillette, 2023; Kendal et al., 2009; Laland, 2004; Laland & Plotkin, 1990). Empirical research on how multiple animals acquire information and change performance simultaneously in a group is limited (Biro et al., 2016).

Over the past decade, however, a small number of collective learning models and mechanisms have been proposed (Collet et al., 2023; Falcón-Cortés et al., 2019; Gildea et al., 2025; Kao et al., 2014). These models make testable predictions of animal performance in learning situations. Consider, for instance, the Conspecific Cue Model (CCM) put forth by Gildea et al. (2025) to explain how collective learning arises from individuals forming simple associations between the actions of their peers and relevant outcomes. A defining feature of CCM is that individuals not only learn from these associations but also influence others through their own actions, leading to collective behavioral shifts. The CCM consists of two key components: (1) a learning mechanism where reward expectations are updated based on observed cues and feedback following a simplified version of the Rescorla-Wagner rule (Rescorla & Wagner, 1972); and (2) an action-selection process where these expectations guide behaviors that, in turn, become cues for others. In CCM, communication between individuals is embedded within the learning process, as the actions of conspecifics provide reward-related information (cf. Gildea et al., 2025).

Consider a simple learning task in which individual and collective performance may be contrasted: the reversal learning task (Sanabria et al., 2024). In the acquisition phase of this task, a simple stimulus discrimination is reinforced—for instance, response A is reinforced in the presence of stimulus A’ and response B is reinforced in the presence of stimulus B’—until it is acquired. In the reversal phase, reinforcement contingencies are reversed—response B is reinforced in the presence of stimulus A’ and response A is reinforced in the presence of stimulus B’. The CCM predicts that training subjects in groups rather than individually yields superior performance in the reversal phase but not the acquisition phase, and that this advantage transfers to the individuals trained collectively, even when tested in isolation. According to this model, the superior performance of collectively trained subjects during reversal emerges from the discriminative function that the behavior of other subjects acquires over the course of training. That is, the CCM predicts that, in an otherwise novel but stable environment (i.e., during acquisition), conspecifics likely constitute unreliable cues that do not contribute to learning; however, when the meaning of these cues becomes ambiguous (i.e., during reversal), the same conspecifics enhance learning by providing informative guidance that is maintained when tested in isolation.

Given the dearth of experimental research on collective learning, the present study seeks to provide a first empirical exploration on the impact of individual versus collective training in the rat—a laboratory animal model that is both conventional and gregarious—using two reversal learning paradigms to test predictions of the CCM. The first paradigm implements reversal learning using a simultaneous discrimination task (Experiment 1); the second paradigm implements reversal learning using a serial (go/no-go) discrimination task (Experiment 2).

Experiment 1

Methods

Subjects

Sixteen experimentally naive female Wistar rats (Rattus norvegicus marked R030-37 and R042-49; Charles River Laboratories, Hollister, CA) served as subjects. They were approximately 65 days old and weighed approximately 200 g at the beginning of the study. Throughout the experiment, rats were housed in groups of four in a home cage consisting of 54 × 28 × 22 cm (L x W x H) transparent plastic box with a wire roof, a layer of Sani-Chip bedding, a piece of PVC pipe, and shredded paper. Home cages were stored in a single colony room housing only female rats on a reverse 12:12 h light/dark cycle kept at a temperature of roughly 22 ℃. The rats were food restricted (1 h of ad libitum feeding starting an average of 50 min after completing the day’s training session) and maintained at 85 ± 5% ad libitum weight. The rats were never fed alone. Depending on their training condition, all four were fed together in their home cage or with a single cage mate in a separate 45 × 22 × 22 cm (L x W x H) transparent plastic cage.

Apparatus

The equipment consisted of a custom-made operant arena measuring 65 × 67.5 × 44 cm (L x W x H) designed for testing multiple rats simultaneously (Fig. 1). The side and back walls were made of opaque black acrylic, the front wall of aluminum, and the floor consisted of a thin metal grate suspended above two catch trays. The entire apparatus was separated into two chambers by two opaque acrylic walls with an opening to allow passage between the chambers. The passage was a 14 cm wide opening set 8.5 cm from the back wall. A removable clear plexiglass sheet was placed on top of the chamber to prevent rats from escaping. The front wall in each chamber consisted of four 5 × 5 cm magazine openings raised 2.5 cm above the floor grate that provided access to a dipper (ENV-202 M-S) separated from each other by a 1 × 1 × 39 cm wall. Each dipper was attached with a 0.01-cc cup (ENV-202C) that delivered the reinforcer [1:3 v/v solution of a Walmart (Bentonville, AR) sweetened condensed milk:water mixture]. The magazine entry was outfitted with a head-entry detector (ENV-254-CB). Two house lights (ENV-215 M-LED) were placed 7.5 cm directly above the two center magazine openings in each chamber. A multiple tone generator (ENV-223) raised 4.5 cm above the floor grate was centered on the outside of each side wall (not used in Experiment 1). All inputs and outputs were programmed and recorded on a computer using Med-PC V software (Med Associates, Fairfax, VT).

Fig. 1.

Experimental Arena. Note. Diagram of arena used for training and testing. The back wall, one side wall, and the plexiglass cover are removed for visibility. Dimensions are to approximate scale

Procedure

Experimental sessions were conducted 7 days/week in a separate experimental room illuminated only by a dim red light and kept at 22°C. Sessions were conducted at about the same time of day for each rat and in the same order across rats, with small variations between training and testing sessions, but always during the dark phase of the light cycle.

The rats in each home cage were assigned, altogether, to one of two training conditions—individual and collective (Table 1). All rats were first magazine trained; the experiment proper consisted of an acquisition phase followed by a reversal phase. In the acquisition phase, one of two stimuli—steady light or flashing (0.25 s on/off) light—indicated the availability of the reinforcer (S +); the other stimulus indicated the absence of the reinforcer (S-). In the reversal phase, the assignment of stimuli as S + and S- were reversed. At the end of each phase, learning was assessed under two test conditions, similar to individual and collective training sessions, but conducted under extinction.

Table 1.

Training Condition and Stimulus Assignment in Experiment 1

Group	Training Condition	Subjects	Acquisition Phase S +	Reversal Phase S +
1	Collective	R030-33	Steady Light	Flashing Light
2	Collective	R034-37	Flashing Light	Steady Light
3	Individual	R042-45	Steady Light	Flashing Light
4	Individual	R046-49	Flashing Light	Steady Light

S + Reinforced conditioned stimulus

Magazine Training

Rats were placed individually in the arena and given a 5-min acclimation period, during which they could explore both arena chambers. Training trials started after this period. On each trial, all eight dippers were then turned on (i.e., the reinforcer made accessible); 3 s after detecting a head entry, the corresponding dipper was turned off. Ten s after lowering all dippers, they were all activated again, indicating the beginning of a new trial. This cycle continued until the rat had collected 40 reinforcers (5 trials). Upon the session's completion, rats were returned to their home cage.

Acquisition

 After all rats successfully completed magazine training, they completed a series of daily training sessions followed by two acquisition tests. The training and testing procedures were the same across individual and collective conditions, except for the number of rats placed simultaneously in the arena: only one rat in the individual condition and four co-housed rats in the collective condition (Fig. 2).

Fig. 2.

Design of Experiments 1 and 2. Note. Visualization of the design for Experiments 1 and 2. X and Y refer to the two stimuli in each experiment with the + and – indicating which stimulus was paired with reinforcement in each phase. For Experiment 1, acquisition training lasted between 26 and 40 session and exactly 40 sessions for reversal training. For Experiment 2, acquisition training ranged from 11 to 42 session and 17 to 42 for reversal training. In both experiments, order of test context, collective and individual, was counterbalanced to control for order effects

Daily training sessions lasted roughly 30 min, consisting of a 3-min acclimation period and 80 trials. Each trial began with a 3-s intertrial interval (ITI), during which no stimuli were presented, followed by a simultaneous presentation of the S + in one chamber of the arena and the S- in the other for 15 s. The presentation of the stimuli was pseudo-randomized (sampled from a list of 20 items), with each stimulus presented on the left and right side of the arena on 10 trials each. Immediately following the cessation of the stimulus, all four dippers were turned on in the chamber corresponding to the S + for 2 s only. An observer manually scored each trial for each rat during the session, with a correct response defined as the rat’s head (nose to front legs) being completely on the side of the S + upon its termination. After each session, individual scores were computed as the proportion of correct responses. The proportion of correct responses was computed as the number of correct responses divided by the sum of correct and incorrect responses. Training sessions were conducted for each group (Table 1) for at least 10 sessions and until one of two criteria were met: (1) each group’s mean proportion correct was 0.75 or more in three out of five consecutive sessions, and no more than one rat scored below 0.65; or (2) 40 sessions were conducted.

Following training, every rat underwent two test sessions, one individual and one collective, implemented in counterbalanced order within training condition (Fig. 2). Testing took place over 3 consecutive days. Half of the rats in each condition completed an individual test on day 1 with the other half left undisturbed in their home cages. Next, all rats completed a collective test on day 2, with the remaining individual test completed on day 3. Individual test sessions were conducted on different days to counterbalance the order of test conditions across individual rats. Each test session consisted of five trials organized similarly to individual or collective training sessions, with the same programmed events occurring on both chambers of the arena, but without reinforcers delivered after any trial. To minimize bias, test sessions were video recorded and independently scored by two observers. Performance was calculated in the same way as training sessions. Following the last acquisition test session, an additional training session was conducted to reinstate food seeking prior to the reversal phase to minimize testing carryover effects.

Reversal

 This phase was organized similarly to the acquisition phase in all respects, except that the assignment of stimuli to S + and S- was reversed (Fig. 2).

Data Analysis

To estimate differences in learning rate between the individual and collective training conditions, a learning curve was fit to each rat’s performance across acquisition and reversal training sessions. The learning curve assumed that the proportion of correct responses, y, increased on every training session proportionally to its differences from unity, y(x + 1) = y(x) + β * [1 – y(x)], where β is the learning rate and x indexes the session. This simple learning rule is consistent with Bush and Mosteller’s (1953) model. Solving it for any session x yields

$$y\left(x\right)=1-\left[1,-,y,\left(1\right)\right]\ast {\left(1,-,\beta \right)}^{x-1}$$ 1

For analysis purposes, estimates of learning rate β were transformed into the number of training session to reduce incorrect response by half, h, halfway between y(1) and unity [e.g., the number of sessions expected for y = 0.75 if y(1) = 0.50],

$$h=1-\frac{1}{{\text{log}}_2\left(1,-,\beta \right)}.$$ 2

Equation 1 was fit separately for acquisition [y_A(x)] and reversal [y_R(x)], with the assumption that y_A(1) = 0.5, that is, that there was no bias toward one CS or another at the onset of acquisition training. Thus, separate halfway sessions, h_A and h_R, were estimated for each rat from fitting these equations to performance in the corresponding training phases:

$$y_A\left(x\right)=1-0.5\ast {\left(2,-,2^{\frac{1}{h_A-1}}\right)}^{x-1},$$ 3

$$y_R\left(x\right)=1-\left[1,-,y_R,\left(1\right)\right]\ast {\left(2,-,2^{\frac{1}{h_R-1}}\right)}^{x-1}.$$ 4

Because session counter x resets at the onset of the reversal phase, y_R(1) is a free parameter representing the proportion of correct (previously incorrect) responses at the onset of the reversal phase. Curve fitting was conducted using the method of least squares.

To determine the effect of training condition (collective vs. individual), training phase (acquisition vs. reversal), and acquisition stimulus (steady vs. flashing light) on training performance, a 2 × 2 × 2 mixed-design ANOVA was conducted on ln(h) and a 2 × 2 ANOVA on logit[y_R(1)]. To determine the effect of these independent variables and test condition (collective vs. individual) on test performance, a 2 × 2 × 2 × 2 mixed-design ANOVA was conducted on logit[y (test)]. Effects were deemed significant with a criterion of  α = 0.05 and were followed up with corresponding lower-order ANOVAs and t-tests.

Results

Although there was significant variability in the goodness of fits among rats, Eqs. 3 and 4 fit each rat’s training performance moderately well, explaining 44.6% (SD = 20.7) of the variance in individual performance across acquisition sessions and 37.5% (SD = 20.9) across reversal sessions (Figs. 3 and 4). Individual parameter estimates for both conditions varied considerably, with the estimated number of training session to reduce incorrect response by half (h) ranging between 23 and 460 for acquisition and 35 and 232 for reversal (Figs. 3 and 4). For training sessions, results of the 2 × 2 × 2 ANOVA revealed no significant main effect of condition on learning rate [F(1,12) = 1.290, p = 0.278]. There was, however, a significant main effect of both phase and stimulus on learning rate, with animals in both conditions learning faster on acquisition sessions than on reversal sessions [F(1,12) = 8.727, p = 0.012] and when reinforcement was paired with the flashing light [F(1,12) = 7.336, p = 0.019]. In addition, there was a significant phase × stimulus interaction effect [F(1,12) = 31.613, p =  < 0.001] with rats learning faster during acquisition training when the S + was a flashing light compared to a steady light S + , t(7) = 4.210, p_holm = 0.006, and compared to reversal training with a flashing light S + , t(7) = 6.065, p_holm =  < 0.001, and with a steady light S + t(7) = 3.570, p_holm = 0.015. Although, during both phases, individual training supported faster learning on average than collective training (Fig. 5), this difference was not statistically significant.

Fig. 3.

Individual Acquisition and Reversal Curves for Collectively Trained Rats in Experiment 1. Note. Curves are fits of Eqs. 3 and 4 to Experiment 1 data. The vertical dotted line separates the acquisition and reversal training phases; the horizontal lines indicate chance performance (.5) and the performance criterion (.75). h = sessions to reduce incorrect responses by half, Y₁ = baseline reversal performance

Fig. 4.

Individual Acquisition and Reversal Curves for Individually Trained Rats in Experiment 1. Note. Curves are fits of Eqs. 3 and 4 to Experiment 1 data. The vertical dotted line separates the acquisition and reversal training phases; the horizontal lines indicate chance performance (.5) and the performance criterion (.75). h = sessions to reduce incorrect responses by half, Y₁ = baseline reversal performance

Fig. 5.

Training Sessions to Reduce Incorrect Responses by Half in Experiment 1. Note. Columns and error bars represent back-transformed means and SEM, respectively

For test sessions, results of the 2 × 2 × 2 × 2 ANOVA revealed no significant main effect of condition [F(1,12) = 3.114, p = 0.103], phase [F(1,12) = 2.643, p = 0.130], context [F(1,112) = 3.233, p = 0.097], or stimulus [F(1,12) = 0.009, p = 0.926], on test performance (Fig. 6). There were also no significant interaction effects between independent variables. Lastly, a 2 × 2 ANOVA on the proportion of correct responses on the onset of reversal training revealed no significant main effect of condition [F(1,12) = 0.010, p = 0.921], or stimulus [F(1,12) = 0.554, p = 0.471]. There was also no significant condition by stimulus interaction [F(1,12) = 0.041, p = 0.842].

Fig. 6.

Experiment 1 Test Session Comparison. Note. Circles and error bars represent back-transformed means and SEM, respectively

Discussion

Previous work on collective learning suggests that rats learn tasks faster in a group than individually (Nagy et al., 2020), which is consistent with predictions from the CCM (Gildea et al., 2025) for reversal performance. Analysis of learning rate for each rat across phases was inconsistent with these previous findings. However, although not statistically significant, the higher learning rate of individually trained rats during acquisition is consistent with CCM predictions (Fig. 5). Nonetheless, CCM predicts faster reversal learning in rats trained collectively rather than individually, which is inconsistent with the present findings. Although individually trained rats show potentially superior training performance relative to collectively trained rats, this difference appears to flip in some test contexts (Fig. 6); this difference, however, is not statistically significant.

The qualitative pattern of results reported in this experiment, although far from conclusive, suggests potentially divergent processes occurring during individual and collective training, perhaps obscured by limitations in the design of the experiment. First, the simultaneous discrimination choice task used in this experiment proved difficult for most rats to learn, with only 2 of the 16 rats meeting the learning criterion of a 0.75 proportion of trials correct in both experimental phases. This may be a result of how the training arena was arranged: an opaque divider was placed between the arena chambers making it difficult for the rats to be simultaneously exposed to both stimuli and to all other conspecifics on each trial, which may have affected the learning dynamics in both conditions. Moreover, the short 2-s window of dipper activation may have limited the rat’s opportunity to access the reinforcer, which may have slowed learning.

Second, our analysis showed a significant stimulus effect with rats learning more quickly with the flashing light S + . Previous research has shown that flashing lights are more effective excitatory than inhibitory discriminative stimuli (Salmon et al., 1988). Thus, the variance in performance induced by the variance in efficacy of the conditioned stimuli may have obscured the differences in learning processes across training conditions (Kerzel & Schönhammer, 2013; Muir, 1996; Rumbaugh et al., 2007).

Third, anecdotal observations suggest that collective training induced considerable play behavior among rats, even though they had continual opportunities to play among them in their home cages. Even in individual training, rats regularly attempted to climb the walls of the arena, occasionally interfering with responding to trials. In addition, although four feeders were provided, observers noted that the rats often competed over access to individual feeders, frequently pushing one another away to gain access to the reinforcer. The maze-like design of the two-chamber arena might have facilitated this behavior, which may have contributed to the slight—albeit nonsignificant—retardation in acquisition and reversal performance (Fig. 5). It is remarkable, nonetheless, that despite this potential interference, collective training yielded test performance that was comparable—when not superior—to individual training.

Finally, the short test sessions, each consisting of just five trials, may not have provided adequate opportunities to obtain sensitive measurements of learning. This can be seen in the substantial variation between animals and between phases with test scores ranging from 0 to 100% correct responding.

Experiment 2

To address some of the limitations of Experiment 1, we conducted a second experiment using a simpler go/no-go task in a single arena chamber. This design ensured that rats were exposed to only one stimulus (S + or S-) per trial, balancing stimulus exposure across sessions. In addition, test sessions in Experiment 2 were matched in duration to training sessions, providing sufficient opportunities for learning to be observed. This experiment was designed to investigate collective learning as an emergent property of group dynamics. Although the model accounts for individual-level learning, data were collected at the group level to capture overarching patterns of learning and interaction within the group.

Methods

Subjects

Sixteen experimentally naive female Sprague Dawley rats (Rattus norvegicus marked R001-8 and R017-24, served as subjects (Charles River). At the beginning of the study, the rats were approximately 70 days old and weighed approximately 180 g. Rats were housed in groups of four in large tubs 100 cm × 80 cm × 35 cm (L x W x H) with a metal roof grate, a layer of Sani-Chip bedding, with access to running wheels and PVC pipes in a single colony room containing only female rats on a reverse 12:12 h light/dark cycle kept at approximately 22 °C. The increase in size and physical enrichment of the home cages relative to Experiment 1 was intended to facilitate play behavior outside the experimental arena, thus potentially reducing this behavior during training and testing. Rats were food restricted (1 h of ad libitum feeding starting an average of 45 min after completing the day’s training session) and maintained at 85 ± 5% ad libitum weight. Rats were fed in a separate 45 × 22 × 22 cm (L x W x H) transparent plastic cage in groups of four.

Apparatus

The equipment used was identical to that in Experiment 1, except that the 14-cm passage between the chambers was blocked by inserting an opaque acrylic inset, creating two separate chambers. All sessions were conducted in only the left chamber.

Procedure

Experimental sessions were conducted daily, 6–7 days/week, in a separate experimental room illuminated only by a dim red light and kept at 22 ℃. Sessions were conducted at about the same time of day for each rat and in the same order across rats, with small variations between training and testing sessions, but always during the dark phase of the light cycle.

The design of Experiment 2 was similar to Experiment 1 (Fig. 2). Rats were randomly assigned to one of two training conditions, individual or collective. As with Experiment 1, experimental sessions involved an acquisition phase and a reversal phase. In the acquisition phase, one of two stimuli—flashing light (0.5 s on/off) or a steady (10-kHz tone)—signaled the availability of the reinforcer (S +); the other stimulus signaled the absence of the reinforcer (S-). The tone was selected to minimize discrepancies in salience between stimuli and to reduce stimulus generalization. In the reversal phase, the stimulus assignment was reversed. Each training phase was followed by two test sessions, an individual test and a collective test, conducted under extinction.

Magazine Training 

All rats completed two sessions of magazine training. Both sessions began with an acclimation period lasting 5 min for the first session and 10 min for the second, during which time they could explore the chamber. Following the acclimation period, eight trials were conducted. At the start of each trial all four dippers were raised. Dippers remained up for 2.5 s following a head entry detection, providing a consistent 2.5-s window to consume each reinforcer. Trials ended once all four dippers were lowered and were followed by a 10-s ITI. This cycle continued until the rat had collected 32 reinforcers.

Acquisition

Once all rats had completed both sessions of magazine training, the experiment continued to the acquisition phase. This phase consisted of a series of daily training sessions followed by two acquisition tests. Training and test sessions were similar between conditions; the only difference was the number of rats simultaneously placed in the arena: one rat for the individual condition, all four group-housed rats for the collective condition.

Each daily acquisition training session lasted 45 min and consisted of a 15-min acclimation period followed by sixty 30-s trials. Trials were broken into 30 S + trials and 30 S- trials, presented in pseudorandom order (sampled from a list of 20 items) throughout the session. All trials began with a 20-s ITI, during which no programmed stimuli were presented, followed by presentation of a 10-s S + or S-. On S + trials, after 5 s of S + onset, the reinforcer was available, with its delivery contingent on a head entry into the magazine. Reinforcers were available for 2.2 s before being lowered, allowing for a maximum of two reinforcers per trial. A response was recorded only when the infrared beam was broken at least once between 4.5 and 5 s into the S + presentation (i.e., 0.5 s prior to food availability). On S- trials, the reinforcer was not available; responses were recorded as in S + trials, but they counted as incorrect. The proportion of correct responses was computed for each group of four rats as the aggregate number of correct responses divided by the sum of correct and incorrect responses. Sessions were conducted daily for each group of four rats for at least 11 sessions and until the group met one of two criteria: (1) the mean proportion of correct responses was at least 0.667 (responding at least twice as much to the S + than to the S- for three of the previous five session; or (2) 42 sessions were conducted.

Following acquisition training, all rats underwent two acquisition test sessions conducted over 6 days. As in Experiment 1, the same programmed events occurred in test trials as in training trials, except for the presentation of the reinforcer. On day 1, half of the rats in each condition completed an individual test with the other half left undisturbed in their home cages; all rats completed a collective test on day 3; the remaining individual tests were conducted on day 5. As in Experiment 1, individual test sessions were conducted on different days to counterbalance the order of test conditions. On days 2, 4, and 6 all rats completed refresher sessions identical to their respective training sessions, to reinstate food seeking prior to the next test or to the reversal phase (Fig. 2). All test sessions were recorded and manually scored following completion.

Reversal Phase 

Reversal training was organized similarly to acquisition training in all respects, except that the assignment of the stimuli to S + and S- was reversed. Reversal tests were conducted identically to acquisition tests but without a Day-6 refresher as no following sessions were conducted following Day-5 tests.

Data Analysis

Unlike Experiment 1, individual training data was not collected in Experiment 2, so the analysis of training data was based on mean correct responses per trial of each group of four rats; no inferential statistics were conducted on these data. Individual test data, however, were collected for each individual rat under both individual and collective control contexts.

Equations 3 and 4 were fit to the averaged data of each group of rats to estimate h and y_R(1) for each group. Test performance was compared between training conditions, test contexts, stimulus assignments (tone vs. flashing light), and training phases using the logit-transformed proportion of correct responses for each rat using a 2 × 2 × 2 × 2 mixed-design ANOVA, as in Experiment 1. Effects were deemed significant with a criterion of  α = 0.05 and were followed up with corresponding lower-order ANOVAs and t-tests.

Results

Equations 3 and 4 fit the data well, accounting, on average, for 54.5% (SD = 27.4) of the variance in acquisition training data across sessions, and for 84.5% (SD = 13.5) of reversal training data (Fig. 7). Unlike Experiment 1, parameter estimates were relatively consistent between conditions, with the estimated number of sessions for learning to reach halfway to asymptote (h) ranging between 7 and 73 for acquisition training and between 16 and 53 for reversal training (Fig. 7). In addition, although not significant, both groups trained with the tone S + learned at a faster rate (17–20 and 21–24, for acquisition and 1–4 and 5–8 for reversal) than those trained with a flashing light S + .

Fig. 7.

Averaged Acquisition and Reversal Curves for Groups of Rats in Experiment 2. Note. Curves are fits of Eqs. 3 and 4 to Experiment 2 data. The vertical dotted line separates the acquisition and reversal training phases after session 42 with the horizontal linen indicating chance performance (.500) and the learning criteria (.667). h = sessions to reduce incorrect responses by half, y_R(1) = baseline performance

No significant main effects of condition [F(1,12) = 0.890, p = 0.364], phase [F(1,12) = 0.005, p = 0.946], or stimulus [F(1,12) = 0.319, p = 0.583] were detected on test performance. There was, however, a significant main effect of test context [F(1,12) = 9.842, p = 0.009], with rats tested individually performing better than those tested collectively across both phases. There was also a significant training condition × test context interaction effect [F(1,12) = 7.581, p = 0.017], with individually trained rats performing significantly worse when tested collectively than when tested individually, t(14) = 4.165, p_holm = 0.007, and worse than collectively trained rats when tested collectively, t(14) = 5.095, p_holm = 0.002 (Fig. 8).

Fig. 8.

Test Performance in Experiment 2. Note. Circles and error bars represent back-transformed means and SEM, respectively. * p < .05

Discussion

Consistent with the results of Experiment 1, individually trained rats learned on average faster than collectively trained rats during the acquisition phase. Although there is substantial stimulus-driven variance, with rats learning the tone S + more quickly than the light S + discrimination, the overall individual training advantage during acquisition is, again, consistent with previous empirical results (Nagy et al., 2020) and with CCM predictions (Gildea et al., 2025). However, in contrast with Experiment 1, collectively trained rats learned, on average, faster than individually trained rats during the reversal phase. These findings are also consistent with CCM predictions. Although subject-specific training data could offer additional detail, our findings remain robust within the framework of group-level learning dynamics. Moreover, whereas the test performance of collectively trained rats was not context dependent, the test performance of individually trained rats was: individually trained rats performed better when tested individually than when tested collectively. This result suggests a protective effect of collective training that supports test performance that is robust to changes in context. It is also possible that the novel presence of conspecifics interferes with test performance substantially more than the novel imposition of isolation.

Conclusions

Despite substantial variability in learning rates—mostly driven by stimulus effects—the present study provides informative guidance for future research on collective learning. In general, when learning a cue-outcome association, the co-presence of socially relevant stimuli, such as cage mates, appears to compromise learning rates during training and testing performance in rats. This effect may be related to play behavior and other competing social activities. Nonetheless, this interference does not appear to undermine associative learning. When tested individually, rats that were trained collectively performed at levels that were at least comparable to rats that were trained individually (Figs. 6 and 8). Moreover, in a go/no-go preparation (Experiment 2), collective training appears to protect test performance from social interference, perhaps because collective training habituates disruptive social responses.

These findings highlight two important considerations for future experimental research on collective learning in animal models. First, stimuli should be carefully selected regarding their salience and other factors that may influence the rate of learning. Second, collective learning must not be confused with collective performance, as learning and performance are dissociable phenomena. In fact, conditions that produce worse performance (e.g., more errors) often lead to better long-term learning outcomes (Soderstrom & Bjork, 2015). To assess the predictions of collective learning models on learning facilitation or interference, learning curves should be drawn not from training performance, but from extinction tests implemented under carefully considered conditions at after various stages of training. Thus far, the evidence suggests some facilitative effects of collective training on learning, but the experimental research on collective learning is still too incipient to draw firmer conclusions.

Supplementary Information

Below is the link to the electronic supplementary material.

Author Contribution

Matthew Gildea: conceptualization, methodology, software, validation, formal analysis, investigation, data curation, writing original draft, writing review and editing, visualization, project administration; Cristina Santos: conceptualization, methodology, software, validation, investigation, data curation, writing review and editing, project administration; Carter Bower: conceptualization, methodology, investigation, writing review and editing; Adeline Hibshman: investigation, writing review and editing, visualization; Federico Sanabria: conceptualization, methodology, validation, formal analysis, resources, writing review and editing, supervision, project administration, funding acquisition; Takao Sasaki: conceptualization, writing review and editing, supervision, funding acquisition.

Funding

This research was supported by a grant from NSF to T. Sasaki and F. Sanabria (award # 2118012).

Data Availability

The data used for the analyses presented in this paper are attached as supplementary material.

Compliance with Ethical Standards

Ethics

All procedures in this article were approved by Arizona State University’s Institutional Animal Care and Use Committee (IACUC), Protocol # 20-1749R and 23-2000R.

Competing Interests

We have no competing interests to declare that are relevant to the content of this article.

Clinical Trial

Clinical trial number: not applicable.