Source memory and collaborative learning: the roles of conflicting information, group composition, and learning partner expertise

Introduction: After collaborative learning, remembering which learning partner shared certain information (source memory) supports social learning strategies such as academic help-seeking: remembering the source helps to ask the right peer when further help is needed.

Methods: A pseudo-collaborative experiment (128 participants) investigated how conflicting information (with conflict vs. without conflict), group composition (heterogeneous vs. homogeneous knowledge levels), and learning partner expertise regarding the learning topic (high vs. medium vs. low) affect source memory and content learning. Multinomial processing tree models estimated source memory unconfounded by guessing.

Results: The mere presence of conflicting information did not affect content learning, but in conditions with conflicting information, those participants who experienced stronger cognitive conflicts learned the content better. Moreover, source memory was better in contexts without conflicting information. Group composition and learning partner expertise did not influence content learning but source memory: in heterogeneous groups, participants remembered the sources better, particularly the learning partners with high expertise.

Discussion: These findings link educational research with cognitive psychology by showing how social and informational factors jointly shape memory processes in collaborative contexts. Considering effects of content- and person-related factors on source memory (and thereby students' help-seeking processes) offers insights into the long-term impact of instructional design and educational tools, ultimately supporting learners in effectively utilizing their peers' knowledge.

1 Introduction

In many educational settings, learning does not happen in isolation. Collaborative learning, defined as “a situation in which two or more people learn or attempt to learn something together” (Dillenbourg, 1999, p. 1), is widely used to foster knowledge acquisition. Thus, students often engage in group-based learning where each member contributes information on a given topic. Collaboration can lead to higher individual learning achievements through various mechanisms, such as (1) deeper engagement with the learning content when confronted with conflicting ideas that require consensus-building (Johnson and Johnson, 2009), which can depend on the presence or absence of conflicting information; and (2) using peers as sources of help by asking questions and providing explanations (Webb, 1989), which can depend on group composition regarding learning partners' expertise. However, while research often focuses on the effects of interventions on content learning or collaborative processes, source memory (i.e., memory for the origin of information, Johnson et al., 1993)—a component of episodic memory—can become relevant for the long-term success of (collaborative) learners. Here, the distinction between content and source becomes relevant: while content learning refers to the to-be-learned subject and material (e.g., facts or concepts), source memory refers to the origin or the context of this information. For example, remembering that the dinosaurs went extinct 66 million years ago is an example of content learning, while remembering that your nephew Alex told you about that is an example of source memory.

Within learning contexts, sources can be multifaceted. While they can refer to artifacts such as documents and remembering them can help to integrate and evaluate different perspectives (Perfetti et al., 1999), in collaborative learning scenarios, information is additionally provided by social sources (i.e., learning partners). For instance, teachers might form small groups with distributed expertise (e.g., Oshima et al., 2017) to facilitate positive (resource) interdependence and thus the exchange of different pieces of information as well as collaborative knowledge construction. When learners remember a piece of information after collaboration, accurately remembering the source in addition can be beneficial for two reasons: (1) facilitating academic help-seeking and (2) informing about the credibility of information in retrospect.

First, academic help-seeking is a self-regulatory social strategy that involves seeking support from peers (Karabenick and Berger, 2013) and can improve academic achievement (Martín-Arbós et al., 2021). This can depend on the awareness of peers' knowledge (Schlusche et al., 2021) and according to the framework of Makara and Karabenick (2013), deciding whom to ask is a crucial aspect of (successful) help-seeking. Knowing where to find information (Keller and Tergan, 2005) and accurately judging who might be the best person to ask (Newman, 2002) are essential self-regulated social competencies. However, students sometimes struggle when deciding which social learning strategy might be the best to apply (Schlusche et al., 2024) and sometimes (unintentionally) rely on inefficient sources of help (Giblin and Stefaniak, 2021): only if learners remember who provided which information can they seek additional information from the most suitable person and select appropriate individuals as future learning partners—highlighting the importance of (source) memory in help-seeking situations.

Second, source memory can inform about the credibility of information in retrospect. For example, a learner could have collaborated with learning partners differing in their expertise regarding the learning topic. While collaborating, one can judge the credibility of presented information based on the expertise of the partner who presented it (Pornpitakpan, 2004). However, if a learner remembers a piece of information several days after the collaboration and wants to judge its credibility, it might be possible that the source is not remembered and thus its credibility cannot be judged accurately. To accurately assess its credibility, learners need to remember whether this information was shared by a peer with high expertise regarding the learning content (indicating that the information might be reliable) or limited knowledge (implying potential unreliability). Forgetting or confusing the sources of information can have grave consequences and false memories might occur (Schacter, 1999). Believing that a piece of information was presented by a partner with high expertise while it was actually presented by a less knowledgeable partner might carry the risk of spreading misinformation and acquiring misconceptions.

For accurate source memory, learners must first encode who contributed which information during collaboration. Here, (cognitive) group awareness plays an important role, which can include learners' awareness of their learning partners' knowledge during collaborative learning (Janssen and Bodemer, 2013) and can contribute to the encoding of social (source) information. Such awareness can be supported with group awareness tools, which collect, transform, and present learning partner and knowledge related information to the learning group (for a review, see Bodemer et al., 2018). For example, some tools provide learners with a knowledge pretest and present the results of learners' performance to the group (e.g., Sangin et al., 2011). Thus, group awareness tools can support learners in building awareness and facilitating collaborative processes, such as asking questions and providing explanations (Dehler et al., 2011; Dehler Zufferey et al., 2010). In doing so, these tools also facilitate the encoding of source-related information by making partner contributions and knowledge levels salient during learning interactions. In the context of collaborative learning, group awareness and source memory both relate to mental representations of links between learning content and learning partners. Group awareness typically reflects a bidirectional association between persons and information maintained in working memory (Schnaubert and Bodemer, 2022), enabling learners to track who knows or has contributed what during interaction. Related constructs in long-term memory include partner modeling and source memory: while partner modeling involves long-term person-information associations from a certain perspective (remembering persons first and inferring information about them), source memory can be conceptualized as similar person-information associations from a different perspective (remembering information first and inferring the associated person).

Research on (computer-supported) collaborative learning has explored effects of person-related factors (such as group composition) or content-related factors (such as conflicting information) regarding their influence on content learning and collaborative processes. However, it is still unclear how such factors influence source memory, and thereby, retrospective learning strategies such as academic help-seeking after collaboration. Analyzing the effects of such factors on source memory can help to derive holistic instructional design principles for collaborative learning. In the following sections, we further focus on conflicting information and group composition as two factors affecting content learning and potentially source memory.

1.1 Conflicting information

When dealing with scientific questions, learners may encounter scientific controversies (i.e., incompatible information about a topic) or conflicting information (information that is inconsistent with prior beliefs or knowledge). Such diverging perspectives and conflicting information can trigger cognitive conflicts and enhance learning. For instance, Maier and Richter (2013) have shown that memory for information from belief-inconsistent texts is better than memory for information from belief-consistent texts, showing that information that conflicts prior beliefs is better learned. In collaborative learning, instructors can deliberately introduce divergent perspectives (e.g., Johnson and Johnson, 2009). When conflicting opinions are presented in a social context such as collaborative learning, these cognitive conflicts are socio-cognitive conflicts in nature and can also be beneficial for learning (Bell et al., 1985; Mugny and Doise, 1978) because learners often try to resolve a cognitive disequilibrium by restructuring cognitions and engage in elaboration processes (Perret-Clermont, 2022).

Regarding source memory, when learners deal with conflicting claims regarding a topic, they need to consider the sources: for example, to avoid misinformation, it can become important to remember whether certain information was presented by a professor working at a public university, who might be perceived as highly credible, or a researcher in industry, who might be perceived as less credible (Thomm and Bromme, 2016). Various frameworks explain how learners attended to and remember sources in contexts with conflicting information (for reviews, see Braasch and Scharrer, 2020; Bråten and Braasch, 2018). Based on the Plausibility-Induced Source Focusing assumption (de Pereyra et al., 2014), readers pay more attention to sources when presented information contradicts prior knowledge. Maier and Richter (2013) also reported better source memory for texts with belief-inconsistent information than for texts with belief-consistent information. However, it is unclear how these effects found in individual (learning) scenarios can be transferred to the context of collaborative learning. Comparing contexts with and without conflicting information can be insightful because while some scientific (learning) topics involve differing perspectives and information (e.g., debates about nuclear power as an energy source), others may naturally often present more consistent and less controversial information (e.g., fundamentals of physics or mathematics).

1.2 Group composition

Learning groups can differ in their composition and groups can be heterogeneous or homogeneous regarding different aspects. Heterogeneous groups where learners differ in their prior knowledge levels sometimes outperform homogeneous groups, especially benefiting learners with lower prior knowledge (Zamani, 2016; Zambrano et al., 2019; Zhang et al., 2015). Such groups offer valuable opportunities for useful interactions: for example, low-knowledge partners might ask more questions to fill their knowledge gaps. These questioning (Palinscar and Brown, 1984) and feedback-seeking (Crommelinck and Anseel, 2013) processes are associated with useful monitoring and learning processes. High-knowledge partners, on the other hand, are then required to elaborate and provide answers, which supports deeper learning (Ploetzner et al., 1999; Webb, 1989). Understanding the effects of heterogeneous and homogeneous groups on content learning and source memory is crucial, as different settings often feature one or the other: while instructors might form heterogeneous groups based on students' prior knowledge (for example, using automatic grouping techniques, for an overview, see Odo et al., 2019), learners themselves tend to self-organize into groups with homogeneous knowledge levels in informal or self-directed environments (Razmerita and Brun, 2011).

Regarding source memory, research suggests that sources are better remembered when they are perceptually more distinct from each other (Bayen et al., 1996) or differ regarding their expertise or trustworthiness (Thomm and Bromme, 2016). Consequently, in heterogeneous groups, where learners differ in expertise, source memory is likely to be better than in homogeneous groups with less variation. Moreover, source memory seems to be better in contexts where one of the sources is untrustworthy (Bell et al., 2022) or with higher rates of misinformation (Pena et al., 2017). Thus, the presence of a low-knowledge learning partner in heterogeneous groups, who may be perceived as less trustworthy due to lower expertise (Pornpitakpan, 2004), could also lead to higher source memory compared to homogeneous groups (with only medium- or high-knowledge partners).

1.3 Source memory for different knowledge levels

When disentangling source memory in heterogeneous groups (e.g., when learning with learning partners who have high, medium, and low knowledge levels), source memory for the individual learning partners might also differ. In the present study, we focus on the role of knowledge levels as an indicator of a learning partner's expertise, which represents one key facet of source credibility, alongside trustworthiness (McGrew et al., 2024; Pornpitakpan, 2004). Previous studies have shown that source memory depends on source credibility: for example, Nadarevic and Erdfelder (2013) have shown that source memory was better for high and low credible persons than for those with uncertain credibility, while high and low credible sources were equally well remembered. In their study, sources were manipulated regarding their trustworthiness and they were telling the truth or lying. The present study investigates how these findings apply in a collaborative learning context, with expertise-related (instead of trustworthiness-related) information about learning partners. Different facets of source memory seem to be enhanced when remembering the source is beneficial in the given context (Kroneisen, 2024; Nadarevic and Erdfelder, 2019). Therefore, in collaborative learning, remembering high-knowledge partners (indicating that a piece of information might be reliable) and low-knowledge partners (indicating unreliability) should be better than remembering medium-knowledge partners as sources.

1.4 Investigating content learning and source memory in an experimental study

While different studies have examined the effects of conflicting information, group composition, or partner expertise on content learning in collaborative learning, understanding their effects on source memory can lead to holistic and long-term implications for instructional design: for example, while certain interventions may not affect immediate learning gains, such factors may have positive effects on source memory and thus on long-term social strategies such as academic help-seeking strategies or retrospective credibility judgments. Thus, the reported experimental study examines how group composition and conflicting information affect both content learning and source memory. In a classical source-monitoring paradigm (see also Kuhlmann et al., 2021), participants first study many different unique items from a few distinct sources in a learning phase. Subsequently, they complete a test phase with a mix of previously presented and new information. Participants first decide whether the item is old or new (item memory). If judged as old, they then indicate its source (source memory). When memory fails, either judgment may be based on guessing. In the present study, participants first read a base text (with or without conflicting information) and then received additional information from three simulated learning partners (with or without differing knowledge levels). Afterwards, a standard source-monitoring test was applied (to assess source memory) and later a knowledge test (to assess content learning outcome).

Regarding content learning, based on the literature on socio-cognitive conflicts, we expect that learning outcome is higher in contexts with conflicting information compared to contexts without conflicting information (H-L-1). Also, based on literature on group compositions, we expect that learning outcome is higher in heterogeneous groups, where learning partners differ regarding their knowledge levels, than in homogeneous groups, where learning partners have the same knowledge level (H-L-2). We will also exploratively examine the interaction between both factors. Also, in the heterogeneous group conditions, we will examine whether information is learned differently based on the knowledge level of the person presenting it (for example, whether information from high-knowledge partners is learned better than information from low-knowledge partners).

Regarding source memory, based on multiple documents research, we assume that source memory is better in contexts with conflicting information compared to contexts without conflicting information (H-SM-1). Furthermore, based on the idea that better distinguishable sources are remembered better, we assume that source memory is better in heterogeneous groups than in homogeneous groups (H-SM-2). Also, sources are better remembered when doing so provides a benefit (for example, to better judge whether a piece of information is credible or not). Thus, we expect that source memory is better for partners with high knowledge than for partners with medium knowledge (H-SM-3a) and better for partners with low knowledge than for partners with medium knowledge (H-SM-3b).

Finally, explorative analyses include relationships between our main variables and prior knowledge, personal interest regarding the learning topic, and the personality trait Need for Cognitive Closure, which refers to an individual's tendency to seek certainty and avoid ambiguity when processing information (Webster and Kruglanski, 1994). A conceptual overview of experimental factors and expected effects is presented in Figure 1.

Figure 1

Figure 1. Conceptual model linking experimental factors, hypotheses, and dependent variables.

2 Methods

The dataset supporting the conclusions of this article and further supplementary information are available in the OSF repository at https://osf.io/ws4m8/ (further materials are available from the corresponding author upon reasonable request).

2.1 Design and participants

We employed a 2 × 2 (×3) design with the between-subjects factors conflicting information regarding the learning topic (with conflict vs. without conflict) and group composition regarding the knowledge level of the (alleged) learning partners (heterogeneous group vs. homogeneous group). In the heterogeneous group conditions, we additionally analyzed the within-subjects factor partner knowledge level (high vs. medium vs. low) which was absent in the homogeneous condition, where all partners were described as having medium knowledge. Participants first read a base article and later received information from three learning partners before taking source memory and learning tests. Key phases of our experiment are depicted in Figure 2. This study employed a pseudo-collaborative design with simulated learning partners. While this controlled setup limits ecological validity compared to real collaborative settings with genuine interaction, it was chosen to ensure high internal validity and precise control over experimental factors. In particular, consistent distribution of information and presentation across learning partners (sources) were essential to isolate the effects of conflicting information, group composition, and learning partner expertise on content learning and source memory. This aspect is discussed in more detail in the Discussion.

Figure 2

Figure 2. Illustration of the key phases of the experimental procedure.

Initially, N = 138 participants were recruited via online advertisements or on campus at the University of Duisburg-Essen (Germany). Data from ten participants was excluded. One participant had not met the pre-defined criteria of solving at least 5 items in the knowledge acquisition test. We also excluded participants who were older than 40 years due to age-related declines in source memory (Old and Naveh-Benjamin, 2008) and non-native speakers (8 participants), as we aimed to minimize potential comprehension differences in this text-heavy study. Thus, all analyses were conducted with the remaining 128 participants (90 female, 35 male, 3 diverse). The age range was between 17 and 35 years (M = 21.32, SD = 3.11) and most of them were students of the Applied Cognitive and Media Science program (82.03%) at the University of Duisburg-Essen (Germany), who received course credit for research participation. Participants (138) were first randomly assigned to the four conditions of our experiment. After excluding data, the rest of the participants (10) were semi-randomly assigned to the conditions to ensure equal group sizes: after one condition had 32 participants, the remaining participants were randomly assigned to the remaining conditions. The experiment was approved by the local ethics committee, lasted approximately 1 h, and standardized instructions were presented on a computer.

To calculate the needed sample size, we conducted an a-priori power analysis using G^*Power (Faul et al., 2007). Given α = 0.05, a target power of 1 – β = 0.80, and a medium sized effect f = 0.25, we needed at least 128 participants. With our final sample size of 128 participants and 54 items in the source memory test (i.e., N = 6912 observations), we were able to detect small effects of the size w = 0.04, which translates to a difference of Δd = 0.18 in source memory parameters.

2.2 Material and procedure

The learning subject in this experiment covered theories regarding the mass extinction event during the Cretaceous-Paleogene boundary 66 million years ago: the meteorite hypothesis (henceforth dominant hypothesis) and the volcanism hypothesis (henceforth alternative hypothesis). We used texts based on the learning material used in prior studies (e.g., Buder and Bodemer, 2011; Heimbuch and Bodemer, 2017), but adapted them for our study purposes. The comprehensibility of instructions, knowledge test items, and materials for the source memory test were tested in two separate pilot studies (n₁ = 20 and n₂ = 27). After filling out an informed consent form, participants had to fill out the demographic questionnaire, which collected information regarding age, gender, and university course. Afterwards they were instructed to write down, in a couple of sentences, everything they know about the extinction of dinosaurs through an open-ended question.

(1) Base article.

Next, participants were instructed to carefully read a base article about the learning subject, as the information would be relevant in a later phase of the study. The base articles differed depending on the conflicting information condition. Participants in the “with conflict” conditions were informed that the two most popular and controversially discussed hypotheses regarding the extinction of the dinosaurs are the dominant hypothesis and alternative hypothesis. The text was heavily in favor of the alternative hypothesis: it contained only one argument for the dominant hypothesis (which was later invalidated) and two arguments in favor of the alternative hypothesis. Participants in the “without conflict” conditions were only informed about the dominant hypothesis: they received the same argument for the dominant hypothesis, but it was not invalidated. Also, they were not informed of the alternative hypothesis. To maintain comparable text lengths, additional general information about meteorites was included. Both texts were of similar length (with conflict: 1,002 words, without conflict: 1,005 words). To ensure that participants adequately read the text, they could click on the “continue” button after at least 6 min had passed. Participants were provided with a countdown at the top right corner of the screen. After 10 min had passed, they continued automatically.

(2) Rating phase 1.

Afterwards, participants rated how likely they found the dominant hypothesis on a scale from 0 (very unlikely) to 100 (very likely). In the “with conflict” conditions, they additionally rated how likely they found the alternative hypothesis using the same scale.

(3) Learning partner descriptions.

Next, participants were informed that they would receive different information regarding the learning topic which (allegedly) originated from the collaboration of three students who first studied the topic independently and then exchanged information. Here, participants were informed that the learning topic remains controversial among scientists, and that it was not verified whether the students used credible sources. Furthermore, participants were told that to preserve anonymity, the students were represented by unisex names (Alex, Luca, Jona). Because the three names were randomly assigned to the learning partners for each participant, from now on we will refer to them as partner A, B, and C (first, second, and third mentioned partner respectively).

Participants were further informed that the learning partners had completed a validated knowledge test on the learning topic. In the “heterogeneous group” conditions, partner A had a high knowledge level, partner B a medium level, and partner C a low level. In the “homogeneous group” conditions, partners A, B, and C were described as having comparable and medium knowledge. To enhance immersion, participants were instructed to imagine collaboratively learning with these partners, who were visually represented by randomly assigned avatars (see Figure 3). The knowledge level of the learning partners was displayed below the avatar and additionally visualized by a star system (3 stars: high knowledge level, 2 stars: medium knowledge level, 1 star: low knowledge level).

Figure 3

Figure 3. Illustrations of the learning phase (a) and the source memory test phase (b). Note For each participant, the names of the learning partners and the avatars (except for the star ratings) were randomly assigned to the different knowledge levels.

To ensure engagement with the sources, participants were told that at the end of the study, they would have to provide feedback and indicate for each partner whether they would choose them as a learning partner. Participants were informed that they would encounter eight different topics, each with three texts (one per partner). They were instructed to memorize the presented information, as their acquired knowledge would later be tested (however, they were not informed about the upcoming source memory test).

(4) Learning and discussion phase.

Figure 3A depicts a trial of the learning and discussion phase. Participants in every condition received the same 24 texts, which were all arguments supporting the dominant hypothesis. Thus, they aligned with the base article in the “without conflict” conditions, but mostly contradicted the base article in the “with conflict” conditions. The texts covered eight different topics (e.g., “acidification of the oceans”, “place of impact”, “the element iridium”), with three separate texts covering unique information for each topic. The first and the last topic (and thus 6 texts) served as fillers to control for primacy and recency effects and no later source memory or knowledge test item referred to the information presented in these filler texts, leaving 18 main texts of similar length (between 49 and 53 words, M_Words = 50.94, SD_Words = 1.21).

Each text appeared on a single page in a fixed order. Within each trial, the topic name was shown first, followed one second later by the source information (i.e., partners' name, knowledge level, and avatar including stars). After 2 seconds, the text appeared, and a countdown started. Participants could click on the “continue” button after 30 seconds. After clicking on the continue button or after 1 min passed, the screen got cleared and the next trial started. For each participant, partners A, B, and C were randomly assigned to one of the three texts of every topic, ensuring that each partner did not always present the same information.

(5) Rating phase 2.

Next, participants provided several ratings. First, they had to indicate how much attention they paid to different source features (see Online Supplementary material in the OSF repository; not focused on in this paper). Then, participants evaluated how competent their learning partners were perceived on a scale from −3 to 3. Each learning partner was judged individually with their respective knowledge level depicted next to the name, e.g., “I think partner C (low knowledge level) is… ‘(not competent at all) −3'/‘(very competent) 3'”. Additionally, learning partners were judged regarding other aspects (see Online Supplementary material in the OSF repository; not focused on in this paper) and participants provided written feedback for both the group and individual partners.

Again, participants rated how likely they found the dominant hypothesis on a scale from 0 (very unlikely) to 100 (very likely), with their previous rating displayed (e.g., “Previously, you rated the dominant hypothesis as X% likely.”). Participants in the “with conflict” conditions additionally rated the alternative hypothesis in the same manner and were then asked to decide which of both hypotheses they found most likely. Participants in the “without conflict” conditions were asked whether they believe the dominant hypothesis is true or not. Afterwards, participants answered some questions regarding their decision (see Online Supplementary material in the OSF repository; not focused on in this paper).

(6) Source memory test phase.

After that, a (previously unmentioned) source memory test followed. For the source memory test, we created old items (i.e., previously presented items from the learning texts) by paraphrasing two pieces of information from each of the 18 texts, and new items (i.e., information that had not been presented before). To reduce possible differences in old–new recognition between items, we examined the proportion of “new” classifications per item in a pre-test (n = 20). While creating new items, the goal was to create three per topic (e.g. “Acidification of the oceans”) with information not presented before in the learning texts. We revised those that were classified as “new” fewer than 16 times (the average of new-classifications for new items), as these items might have been too close to the original learning texts. For old items, the average number of “new” ratings was 3 times and paraphrases that received more than 4 “new” ratings were revised to ensure adequate recognizability as old items, as they may have differed too much from the original learning texts.¹

Before the source memory test of our main study, participants were informed that they would see several learning topic-related information—some of which were paraphrased versions of information previously presented by the learning partners, while others were new and had not appeared before. Before starting, participants were shown examples of an old piece of information (from the filler texts) and a new piece of information that had not been presented before. Figure 3B illustrates an example trial from the test phase. The test consisted of 54 trials, including 36 previously presented items (2 paraphrased items from each of the 18 texts) and 18 new items (3 new distractor items for each of the 6 topics). Each information was a single sentence (M_Words = 12.04, SD_Words = 1.66) and the 54 items were presented in a random order.

Each trial had the same structure. At the top of the screen, the sentence “Who presented the following information?” remained throughout the test phase. Below, the item was presented which changed after the participant clicked on “continue”. Participants could provide source judgments by selecting one of the four options “Partner A”, “Partner B”, “Partner C”, or “No one”. If unsure, they were encouraged to guess. Partners were again represented by their avatar (including the star rating), their name, and their knowledge level. For each participant, the order of the partners was decided randomly once at the beginning of the test, while the option “no one” always appeared last. A progress bar at the top of the screen indicated the percentage of completed trials. After finishing all 54 trials, participants were asked to describe any strategy they used when assigning sources.

(7) Need for Cognitive Closure.

We measured Need for Cognitive Closure with the Need for Cognitive Closure scale (Schlink and Walther, 2007). The validated German questionnaire consists of 16 statements (e.g. “Generally, I avoid engaging in discussions on ambiguous and controversial topics.”). Participants rated each statement on a 7-point scale ranging from “fully disagree” (−3) to “fully agree” (3) and Cronbach's α was acceptable to good (α = 0.79).

(8) Learning test.

The learning test consisted of 18 multiple-choice questions, each containing one target and three distractors. Participants were informed beforehand that the questions would be based on the texts from the three partners. Thus, six questions were presented about the texts from each of the three partners. The questions were designed in such a way that they could be solved without the information from the base articles. Every question was presented on a single page and in a fixed order. Item difficulty (measured as the proportion of participants answering correctly) ranged from 0.48 to 0.89 (M = 0.67, SD = 0.13). Learning test items were tested in pre-tests and revised accordingly.

(9) Final questions and debriefing.

At the end of the study, participants were asked whether they believed that the information from the learning partners originated from the discussion of real students (“Do you think the statements by Alex, Jona, and Luca come from real students?”) on a scale from 3 (“The information definitely originated from real students”) to −3 (“The information definitely did not originate from real students”), M = −0.10, SD = 1.24. Afterwards, participants indicated their prior knowledge on the topic (“How would you have rated your prior knowledge on the topic of ‘extinction of the dinosaurs' before taking part in the study?”) on a scale from 3 (“I had high prior knowledge”) to −3 (“I had no prior knowledge”), M = −1.12, SD = 1.57. To further validate the self-assessments, we examined the open-ended responses from the previous phase, in which participants were asked to write down everything they knew about the extinction of dinosaurs. Specifically, participants who reported high prior knowledge at the end of the study also provided more information about the extinction of dinosaurs in their open-ended responses, whereas participants who reported low prior knowledge often revealed their missing prior knowledge. Finally, participants reported how interesting they found the topic (“How interesting did you find the topic ‘extinction of the dinosaurs'?”) on a scale from 3 (“I found the topic very interesting”) to −3 (“I did not find the topic interesting at all”), M = 1.34, SD = 1.57. Two 2 × 2 ANOVAs suggested that prior knowledge and interest in the topic did not significantly differ between the experimental conditions, F_{(1, 124)} < 3.53, p > 0.062, ${\eta }_p^2$ < 0.03.

2.3 Statistical analyses

For our content learning-related hypotheses H-L-1 and H-L-2, the dependent variable was learning outcome, which was operationalized as the number of correctly solved items in the learning test (0–18) in phase 8. To test these hypotheses (and exploratively the interaction effect), we calculated a 2 × 2 ANOVA with the factors “conflicting information” (with conflict vs. without conflict) and “group composition” (heterogeneous group vs. homogeneous group). For the explorative question regarding the effects of learning partner expertise on learning, learning outcome is separately calculated for the information from the three learning partners with different knowledge levels (high vs. medium vs. low), thus ranging from 0 to 6.

Regarding source memory, we complementarily used two measures.

• We calculated classification-based measures for separate sources (learning partners). So-called Conditional Source Identification Measures (CSIM, Bröder and Meiser, 2007; Murnane and Bayen, 1996) are calculated by dividing the number of correct source judgments by the sum of “old” responses (i.e., “high”, “medium”, “low” responses in heterogeneous groups or “A”, “B”, “C” responses in homogeneous groups). For more specific examples, see Appendix A. However, such measures can sometimes confound cognitive processes which are involved in source monitoring judgments, such as source memory and several guessing biases (Bröder and Meiser, 2007; Murnane and Bayen, 1996). An example of such a guessing bias would be a strategy like: “If I can't remember the source, I will rather guess that the partner with medium knowledge presented the information”—a strategy which can be applied consciously or subconsciously. Such a bias could distort CSIM scores for the medium-knowledge partner, leading to the incorrect assumption that memory for this partner might be unusually strong.

• To control for guessing processes, we used a multinomial model-based approach (for overviews, see Erdfelder et al., 2009; Singmann et al., 2024). Such mathematical models can be used to disentangle different cognitive processes (e.g., memory and guessing) and measure their probabilities based on observed response frequencies (see Table A1 in Appendix A) in the memory test. The validated two-high threshold multinomial source monitoring model (Bayen et al., 1996) has been used in many previous studies, both with two sources (e.g., Mieth et al., 2021; Symeonidou and Kuhlmann, 2021) and three sources (e.g., Buchner et al., 2009; Keefe et al., 2002; Nadarevic and Erdfelder, 2013). Analyses were conducted using the program multiTree (Moshagen, 2010). An adaptation of the model for our present experiment with three sources is illustrated in Figure 4. Our joint MPT model is described in more detail in Appendix A.

Figure 4

Figure 4. Multinomial processing tree model of source monitoring (Bayen et al., 1996), adapted for three sources (Keefe et al., 2002) and the present study. Note Rectangles on the left side represent the learning partner who presented a piece of information in the learning and discussion phase. In heterogeneous groups: i ε {high, medium, low}. In homogeneous groups: i ε {A, B, C}. Rectangles on the right side represent the possible answers of participants in the source memory test. Letters along the links represent probabilities of certain cognitive processes. D = probability of recognizing that a piece of information was presented before (D_i) or that it was not presented before (D_N). d = probability of remembering the source (learning partner). b = probability of guessing that a piece of information was presented before. a_B/Medium = probability of guessing that a recognized piece of information was presented by partner B/Medium. a_A/High = probability of guessing that a piece of information was presented by partner A/High, given partner B/Medium was not guessed to be the source. g_B/Medium and g_A/High = probabilities of guessing the source, given the piece of information was not recognized as previously presented.

Both measures complement each other: while CSIM scores can confound guessing and memory, MPT models can measure source memory unconfounded by guessing. However, our MPT model-based approach estimates memory on a group level and does not consider individual differences,² while CSIM scores are calculated on an individual level and allow to test GLM-based correlation or regression analyses. Therefore, to combine the strengths of both measures, we will proceed as follows while testing our hypotheses and exploratory analyses:

To test H-SM-1 (i.e., that source memory is better in contexts with conflicting information than without conflicting information), we will use MPT model-based analyses and restrict the source memory parameters d_A/High, d_B/Medium, and d_C/Low between the conditions with and without conflicting information. In the same manner, to test H-SM-2 (i.e., that source memory is better in contexts with heterogeneous groups than with homogeneous groups), we will restrict the source memory parameters between the conditions “heterogeneous group” and “homogeneous group”. In the conditions with heterogeneous groups, to test for potential memory advantages for high-knowledge (H-SM-3a) and low-knowledge (H-SM-3b) partners over medium-knowledge partners, we will impose the restrictions d_High = d_Medium and d_Low = d_Medium on the base model. For the sake of comprehensibility, the indices A, B, and C will be substituted by high, medium, and low (representing the knowledge level) in the heterogeneous groups.

To exploratively test for relationships between source memory and other variables, we will use CSIM as a source memory measure. The findings will be interpreted carefully because CSIM can confound memory and guessing processes (Bröder and Meiser, 2007). To derive careful interpretations, we will analyze guessing parameters with MPT models and test whether they deviate from predicted frequencies. If certain parameters deviate, we will omit corresponding CSIM values from the analyses. For example, if the parameter a_B/Medium deviates from 0.33, this would indicate that CSIM_B/Medium strongly confounds memory and guessing and we would exclude CSIM_B/Medium from the data. Thus, we will use (potentially adjusted) average CSIM scores as a source memory measure to test explorative analyses.

3 Results

The level of significance was set to α = 0.05 for all analyses. Post-hoc tests were conducted with Bonferroni-Holm corrected significance levels (Holm, 1979). Corrected degrees of freedom by Greenhouse-Geisser indicate that the assumption of sphericity was violated.

3.1 Manipulation checks

To assess whether the base articles successfully induced the intended conflict, participants rated the likelihood of the presented hypotheses after reading the articles (see Materials and Procedure, phase 2). Mean ratings are depicted in Table 1. Two t-tests yielded the following results: (1) In the conditions with conflict, participants rated the alternative hypothesis as significantly more likely than the dominant hypothesis, t₍₆₃₎ = 9.56, p < 0.001, d = 1.19. (2) Across the groups with and without conflict, the dominant hypothesis was rated as more likely in the groups without conflict than with conflict, t_(117.74) = 14.58, p < 0.001, d = 2.58 (note that we used Welch's t-test because homogeneity of variances was violated). Both results indicate a successful manipulation of conflicting information through the base articles.

Table 1

Table 1. Mean likelihood ratings (in percent, with standard deviations in parentheses) of the dominant hypothesis and alternative hypothesis.

Additionally, two one-way ANOVAs regarding the perceived competence of the learning partners (see Methods, phase 5) revealed that in the heterogeneous groups, competence-ratings differed for the three partners, F_{(1.82, 114.72)} = 25.82, p < 0.001, ${\eta }_p^2$ = 0.29. The high-knowledge partner was perceived as most competent (M = 1.88, SD = 0.86), followed by the medium-knowledge (M = 1.38, SD = 1.02) and the low-knowledge partners (M = 0.89, SD = 1.14). All pairwise comparisons were significant, t₍₆₃₎ > 3.17, p_Holm < 0.002. In the groups with homogeneous learning partners, perceived competence of the partners A, B, and C (first, second, and third mentioned partners, respectively) did not differ, F_{(2, 126)} < 0.01, p > 0.999, ${\eta }_p^2$ < 0.01 (all M = 1.36, SD_A = 1.10, SD_B = 1.09, SD_C = 1.07). As the avatars, names, and texts were randomly assigned to the partners, the observed differences in competence ratings were solely based on the knowledge level label of each partner.

3.2 Learning outcome

Learning outcome (i.e., correctly solved items in the learning test, see Table 2) was analyzed with a two-factorial 2 × 2 ANOVA, with the between-subjects factors conflicting information (with conflict vs. without conflict) and group composition (heterogeneous vs. homogeneous). Contrary to H-L-1, the presence of conflicting information did not lead to better learning than its absence, F_{(1, 124)} = 0.23, p = 0.630, ${\eta }_p^2$ < 0.01. Further, the main effect of group composition was also not significant, F_{(1, 124)} = 2.10, p = 0.150, ${\eta }_p^2$ = 0.02, contradicting H-L-2. Finally, the (exploratively analyzed) interaction between group composition and conflicting information was also not significant, F_{(1, 124)} = 0.50, p = 0.479, ${\eta }_p^2$ < 0.01. In the next steps, we used exploratory analyses to examine nuanced effects of conflicting information and group composition on learning outcome.

Table 2

Table 2. Mean score in the learning test as a function of group composition (heterogeneous vs. homogeneous knowledge levels) and conflicting information (with conflict vs. without conflict).

3.2.1 Cognitive conflict

In the conditions with conflict, not every participant might have perceived the same amount of cognitive conflict. For instance, despite receiving multiple pieces of information from their learning partners supporting the dominant hypothesis (which contradicted the earlier presented alternative hypothesis), 47 participants still considered the alternative hypothesis more likely, while 17 participants favored the dominant hypothesis. To test whether the amount of perceived conflict influenced learning outcome, we decided to conduct two tests.

(a) A t-test within the “with conflict” conditions tested whether learning differed between participants who found one of both hypotheses more likely. Shapiro-Wilk tests indicated no significant deviations from normality in either group, and Levene's test confirmed homogeneity of variances (p > 0.05); therefore, the assumptions for Student's t-test were met despite the unequal group sizes. The two-sided t-test revealed a medium-sized effect, which was, however, not significant, t₍₆₂₎ = 1.98, p = 0.052, d = 0.56. Descriptively, students selecting the alternative hypothesis after the learning and discussion phase had higher learning outcome (M = 12.34, SD = 3.29) compared to those who selected the dominant hypothesis (M = 10.53, SD = 3.06), which the learning test was based on.

(b) Participants in the conditions with conflict rated the likelihood of both hypotheses after the base articles and after the learning and discussion phase. For both phases, we calculated a conflict-value: “alternative hypothesis likelihood minus dominant hypothesis likelihood”. This measure served as an operationalization of socio-cognitive conflict, i.e., the degree to which a participant's current belief diverged from the information presented by their learning partners. Positive values indicate a stronger belief in the alternative hypothesis (which contradicts the information from the learning phase that the learning test is based on) and negative values indicate a stronger support for the dominant hypothesis. After reading the base articles, the mean conflict-value was +36.48% (SD = 30.55%), while this value decreased after the learning and discussion phase (M = +12.50%, SD = 32.12%). Linear regressions revealed that the conflict-value after the base articles did not significantly predict learning outcome, F_{(1, 62)} = 2.52, p = 0.118, R² = 0.04, b = 0.02, β = 0.20. However, the conflict-value after the learning and discussion phase significantly predicted learning outcome, F_{(1, 62)} = 4.63, p = 0.035, R² = 0.07, b = 0.03, β = 0.26. This suggests that the more participants disagreed with the information received from their learning partners, the better they learned the information, indicating that the conflict at the beginning of a collaborative learning session is less determining than the conflict experienced during or after the session.

3.2.2 Partner knowledge level

We further analyzed exploratively whether learning outcome differed depending on the knowledge level of the source (for descriptives, see Figure 5). We therefore conducted a 2 × 3 ANOVA with the between-subjects factor conflicting information (with conflict vs. without conflict) and within-subjects factor partner knowledge level (high vs. medium vs. low) in the experimental conditions with heterogeneous group composition. Learning outcome was measured as the number of correctly solved items in the learning test corresponding to the information presented by each of the three partners (0–6). In line with prior analyses, learning outcome did not differ in contexts with or without conflicting information, F_{(1, 62)} = 0.03, p = 0.877, ${\eta }_p^2$ < 0.01. Also, neither the main effect of partner knowledge level, nor the interaction with the factor conflicting information was significant, F_{(1.90, 117.65)} < 1.63, p > 0.202, ${\eta }_p^2$ < 0.03. Participants learned the information equally well regardless of the competence of the source, both in contexts with and without conflicting information. On average, learning outcome was lower for the information from the partner with a low knowledge level than for the information from the partners with high and medium knowledge levels.

Figure 5

Figure 5. Learning outcome and source memory measures as a function of group composition (heterogeneous vs. homogeneous group), conflicting information (with conflict vs. without conflict), and source of information (heterogeneous group: A/High, B/Medium, C/Low; homogeneous group: A/Medium, B/Medium, C/Medium). High, medium, and low denote the knowledge level of the learning partner who presented the information. A (first), B (second), and C (third) represent the order of learning partners' introduction. Learning outcome: Mean number of correctly solved questions in the knowledge test regarding the information which were originally presented by a certain partner (0–6). Parameter estimate d: Multinomial processing tree model-based probability estimations (0–1) for source memory. Conditional Source Identification Measure: Correct source classifications relative to the number of hits (see Appendix A).

3.3 Source memory

Subsequently, we investigated source memory performance with MPT models. First, we had to define parameter restrictions in order to obtain identifiability of our joint MPT model. The two-high threshold model implies that old–new recognition (D) does not differ between old and new items (Bayen et al., 1996; Snodgrass and Corwin, 1988). To obtain a parsimonious base model, we decided to set all parameters representing old–new recognition to be equal (D_A/High = D_B/Medium = D_C/Low = D_N), within and across all experimental conditions (for a similar restriction, see Kroneisen, 2018). We also set the corresponding a and g parameters for source guessing within each experimental condition equal (a_A/High = g_A/High, a_B/Medium = g_B/Medium). The base model fit the data well, G²₍₂₇₎ = 26.62, p = 0.272, and allowed the old–new recognition parameters (D) to be equated across experimental groups. This indicates that item recognition—potentially a measure of content learning—was not influenced by conflicting information, group composition, or the source of information, further supporting the findings of the learning test. Our hypotheses are based on the source memory parameters d and descriptive data is depicted in Figure 5. For the sake of completeness, the parameter estimates of all the parameters based on our base model are depicted in Table A2 (Appendix A).

To test the first hypothesis H-SM-1 that sources are better remembered in contexts with conflicting information than without conflicting information, we restricted the source memory parameters between contexts with and without conflicting information. There was a significant difference between both conditions, ΔG²₍₆₎ = 14.79, p = 0.022, w = 0.05. Contrary to our hypothesis, however, source memory was better in contexts without conflicting information. Furthermore, the hypothesis that learning partners should be remembered better as sources when they differ regarding their knowledge level (H-SM-2) was supported by our data, ΔG²₍₂₎ = 29.75, p < 0.001, w = 0.07. Next, to test whether source memory was better for the partner with high knowledge than medium knowledge (H-SM-3a), the additional restriction d_High = d_Medium was imposed on the base model. Source memory for high-knowledge partners was better, ΔG²₍₂₎ = 12.22, p = 0.002, w = 0.06. Testing the hypothesis that low-knowledge sources are better remembered than medium-knowledge sources (H-SM-3b) by imposing the restriction d_Low = d_Medium, however, barely missed the level of significance and did not increase model misfit, ΔG²₍₂₎ = 5.86, p = 0.053, w = 0.03. Low-knowledge sources were only descriptively better remembered than medium-knowledge sources.

We further exploratively tested whether sources are remembered equally in homogeneous groups. The partners only differed regarding their order of introduction (A = first, B = second, and C = third) and were described as having medium knowledge. Restricting d_A = d_B = d_C did not increase model misfit, ΔG²₍₃₎ = 3.47, p = 0.280, w = 0.02, indicating that sources are remembered equally well when they do not differ regarding their knowledge level, irrespective of their position of introduction.

3.4 Source guessing and correlative relationships

Before using classification-based Conditional Source Identification Measures (CSIM, for descriptive data, see Figure 5) to test for correlative relationships between source memory and other variables, we tested for potential underlying guessing tendencies with our MPT model-based analyses (for parameter estimates, see Table A2 in the Appendix). In the heterogeneous group conditions, we restricted a_Medium to 0.33 (as all three partners shared the same amount of information), which was significant, ΔG²₍₂₎ = 31.86, p = 0.013, w = 0.07. In the absence of source memory, learners over-guessed that the partner with medium knowledge was the source of information. The parameter a_High represents the tendency to guess that the high-knowledge partner presented a piece of information, given the medium-knowledge partner was not guessed. Thus, we restricted this parameter to 0.50 (as the remaining high- and low-knowledge partners shared the same amount of information), which was not significant, ΔG²₍₂₎ = 0.62, p = 0.733, w = 0.01, indicating no further guessing tendencies between partners with high and low knowledge. In the homogeneous group conditions, restricting a_B to 0.33 did not increase model misfit, ΔG²₍₂₎ = 3.80, p = 0.150, w = 0.02, and restricting a_A to 0.50 was also not significant, ΔG²₍₂₎ = 2.48, p = 0.290, w = 0.02. When all learning partners have the same knowledge level, they seem to be guessed equally often, corresponding to the amount of information they share.

Thus, average classification-based CSIM scores were calculated differently based on the “group composition” conditions: in the heterogeneous group composition conditions, average CSIM was calculated as the mean of CSIM_High and CSIM_Low, as CSIM_Medium confounded memory and guessing (see Section 2.4). In the homogeneous group composition conditions, average CSIM was calculated as the mean score of CSIM_A, CSIM_B, and CSIM_C, since no underlying guessing processes have been found. Note that these average source identification scores are still not process pure measures and only serve as an approximate indicator of source memory. The corresponding mean CSIM scores are depicted in Figure 5.

We used two-tailed Pearson correlation tests to explore correlative relationships. There was no relationship between learning outcome and source memory, r₍₁₂₆₎ = 0.05, p = 0.583. Further relationships between our main dependent variables (learning outcome and source memory) and other variables are depicted in Table 3. Learning outcome correlated positively with prior knowledge and interest regarding the learning content. However, there was no significant relationship with Need for Cognitive Closure. Furthermore, we found no relationship between source memory and any of the other variables.

Table 3

Table 3. Pearson correlation values for exploratory relationships.

Additionally, we explored mean study times as a measure of attentional allocation during the learning phase. These analyses revealed no systematic differences between conditions or relationships with the main dependent variables. Complete results are provided in the Online Supplementary material in the OSF repository.

4 Discussion

In instructional design and (computer-supported) collaborative learning, effects of different rather content-related factors (e.g., the presence or absence of conflicting information) or rather person-related factors (e.g., group composition) are often analyzed with regard to their effectiveness in improving learning or collaborative processes. However, efficient long-term social learning strategies (e.g., academic help-seeking) can rely on long-term memory constructs such as source memory (i.e., remembering the source of information, Johnson et al., 1993). The present study thus investigated the effects of conflicting information and group composition with regard to the knowledge levels of learning partners on content learning and source memory in a pseudo-collaborative learning context.

4.1 Conflicting information

In collaborative learning environments, learners are sometimes confronted with conflicting information, which refers to the presence of inconsistent or contradictory claims presented during learning. Learning outcome was not increased by the presence of conflicting information, which contradicts H-L-1. Potentially, the missing effect could be attributed to the lack of (useful) epistemic conflict regulation between learners, which can be important for the benefits of socio-cognitive conflicts (Buchs and Butera, 2004). However, more nuanced analyses revealed that individual differences in experienced conflict matter. In conditions with conflicting information, conflict-values reflected the level of disagreement between the hypothesis in the base article and the dominant hypothesis presented by the learning partners and used for the knowledge test. Only the (socio-)cognitive conflict measured after collaborative learning predicted content learning: learners who still showed stronger agreement with the alternative hypothesis (i.e., who experienced more conflict) learned more from their learning partners. In contrast, initial conflict based solely on the base articles did not predict learning. The present study underlines a distinction between cognitive conflicts emerging in individual learning contexts and socio-cognitive conflicts in collaborative learning contexts: in individual learning contexts, when learning information from multiple texts, prior beliefs can influence the processing of information and belief-inconsistent information is learned better (e.g., Maier and Richter, 2013). In social settings, when learning with peers as social sources of information, socio-cognitive conflicts that emerge during learning could better explain deeper processing of belief-inconsistent information. However, while socio-cognitive processes may contribute to the observed effect, it may also be driven by the amount of information presented: cognitive conflicts arising from the presence of multiple pieces of information supporting the dominant hypothesis—regardless of whether their sources are social or textual—could likewise account for the findings.

The results regarding source memory further support the idea of a distinction between cognitive conflicts in individual learning (with multiple texts) and socio-cognitive conflicts in collaborative learning. Drawing on literature on the role of conflicts in multiple source use (for reviews, see Braasch and Scharrer, 2020; Bråten and Braasch, 2018), we initially assumed source memory to be better in contexts with conflicting information compared to those without. Contrary to our expectations, source memory was better in a context without conflicting information (contradicting H-SM-1), suggesting that the Plausibility-Induced Source Focusing assumption (de Pereyra et al., 2014) cannot be transferred from individual to collaborative learning contexts. While the presence of conflicting information did not improve source memory in the present study, different collaborative learning contexts could yield different results, as the role of the learner may be different. For instance, in the study of Buder and Bodemer (2008), the authors examined a collaborative learning context in which one learner (the minority) received information predominantly supporting one hypothesis, while three other learners (the majority) received information predominantly supporting another hypothesis. The present study tested source memory and content learning from the perspective of the minority member in the context with conflicting information. However, it remains unanswered how students would learn the content and remember sources when they are part of the majority and some other learners present information supporting the same hypothesis while another learning partner (minority) presents opposing information. While the Plausibility-Induced Source Focusing assumption could not be transferred from individual learning to collaborative learning, it remains untested whether other theoretical accounts such as the Discrepancy-Induced Source Comprehension model (Braasch et al., 2012) could account for source memory in collaborative learning, which proposes that conflicts between texts—and not necessarily conflicts between prior beliefs and texts—can enhance source memory (see also Strømsø et al., 2013). Furthermore, regarding content learning, participants may have learned the information from the three learning partners equally well because all three of them presented information supporting the same scientific hypothesis. Manipulating which learning partner (e.g., the expert or the novice) agrees or disagrees with the learner's claim can help to understand under what circumstances information from learning partners with differing expertise is learned better or worse.

Taken together, the mixed results regarding conflicting information support the idea that socio-cognitive conflicts can improve learning in a collaborative context Bell et al., (1985); Mugny and Doise, (1978); Perret-Clermont, (2022), even without interactions such as consensus-building (Johnson and Johnson, 2009). However, in such contexts with conflicting information, learners struggle more to remember who shared certain information, which might hamper useful long-term social strategies such as academic help-seeking and asking the right peer more questions. Thus, especially in contexts in which instructors make use of structured controversies to help students explore controversial topics, they may consider additional strategies to enhance learners' source memory (which will be discussed in the later Section 4.6).

4.2 Group composition

Learning groups can vary in their composition with respect to multiple learner characteristics. One key dimension is prior knowledge: learners within a group may differ substantially in their knowledge about the topic at hand (heterogeneous groups) or exhibit similar levels of prior knowledge (homogeneous groups). Regarding content learning, the data did not support the hypothesized benefits of heterogeneous groups (H-L-2). Even in a pseudo-collaborative study without interactions between learners, positive effects of heterogeneous groups with different knowledge levels might have been plausible: for example, receiving information from less credible sources can lead to more attention to the content (Wertgen and Richter, 2023). Thus, being aware of differing source expertise (with at least one source with low expertise) can focus learners' attention on the content. While this study examined whether effects of group heterogeneity on learning can occur beyond the collaborative elements themselves, the missing effect in the present pseudo-collaborative study indicates that positive effects of group heterogeneity likely depend on meaningful collaborative processes, such as when less knowledgeable learners seek explanations from more knowledgeable peers (Webb, 1989). However, even in real groups, such strategic communication does not occur automatically (Soller, 2001) and can require guidance. Such guidance can be explicit, for example, through collaboration scripts that structure interactions between learners (Kollar et al., 2006) and have been shown to enhance learning outcome (Vogel et al., 2017). Yet, implicit support can also foster productive engagement: group awareness tools (Bodemer et al., 2018) can guide learners' attention to the high or low knowledge of learning partners, prompting more targeted questions and explanations (Dehler Zufferey et al., 2010).

Source memory, however, was influenced by group composition and was better in heterogeneous groups with different knowledge levels compared to homogeneous groups with same knowledge levels (supporting H-SM-2). This advantage may stem from the increased distinguishability of learning partners: beyond varying names and avatars, participants could also differentiate the sources based on the different knowledge levels, which enhances source memory through better discrimination (Bayen et al., 1996; Symeonidou and Kuhlmann, 2021; Thomm and Bromme, 2016). This study extends prior research on the benefits of heterogeneous groups: students are better able to remember the learning partner who shared certain information in heterogeneous groups, which may facilitate more effective academic help-seeking in retrospect. Note, however, that group heterogeneity does not universally improve source memory: for instance, Pepe et al. (2021) found that participants demonstrated better source memory in homogeneous groups with respect to ethnic diversity. Thus, the positive benefits of group heterogeneity on source memory found in this study may not generalize to all aspects of group composition. Practically, instructors may aim to form heterogeneous groups with differing knowledge levels (for an example of automated grouping in the classroom, see Erkens et al., 2016) to improve collaborative learners' source memory.

4.3 Partner knowledge level

In heterogeneous groups, the specific knowledge level of the learning partner can also play a role in learning and source memory. Regarding content learning, analyses revealed that information from partners with different knowledge levels is learned equally well: for example, learners do not seem to focus more on information from high-knowledge partners and do not seem to disregard information from low-knowledge partners. However, learning information from low-knowledge partners equally well might carry the risk of acquiring misconceptions, as low-knowledge partners may have less accurate or less thoroughly researched information. When learners do not selectively disregard information from low-knowledge partners while learning the content, tagging the source becomes especially relevant: for example, remembering that a certain piece of information came from a low-knowledge partner could prompt them to validate information (Wertgen and Richter, 2023).

Source memory for the three partners differed in heterogeneous groups. High-knowledge partners were better remembered than medium-knowledge partners (supporting H-SM-3a). In contrast, low-knowledge partners were not significantly better remembered than medium-knowledge partners (contradicting H-SM-3b), although the descriptive trend was in the expected direction. These results support the idea that source memory is context-dependent, meaning that how well certain sources are remembered is tied to the perceived value of these sources in the given context (e.g., Kroneisen, (2024); Nadarevic and Erdfelder, (2019). In the present study, remembering that information was presented by a high-knowledge partner seemed more important, as this implied the information was well researched and likely true. This could also be due to students' help-seeking strategies: according to the expectancy-value model of source selection and utilization of Makara and Karabenick (2013), source selection depends on the perceived quality of the source. It is reasonable to assume that a learning partner with high knowledge of the topic is seen as a valuable resource for future help and learners would want to direct questions to them later. Especially with the given task to study for a test and to decide whether to select the students as future learning partners, a context might have been present that guided learners' attention more to high-knowledge peers as sources, but not to their shared content (as the content of all partners was learned equally well). This aligns with the results of Ülker and Bodemer (2025a), who found that source memory for high-knowledge partners is better than for low-knowledge partners only when learners explicitly study for a knowledge test.

4.4 Relationships between learning outcome, source memory, and learner characteristics

To explore relationships between source memory and other variables, we relied on classification-based conditional source identification measures (CSIM). While these measures capture aspects of source memory, they can also include influences of guessing. Consequently, any observed correlations should be interpreted with caution.

Explorative analyses revealed that there was no correlation between learning outcome and source memory. In some contexts, a tradeoff between item memory (here: learning outcome) and source memory can be observed (see, for example Jurica and Shimamura, 1999). However, the evidence in different learning contexts is mixed: while the current experiment found no relationship between learning and source memory, other studies even report positive relationships, in individual learning (e.g., Strømsø et al., (2010) or collaborative learning (Ülker and Bodemer, (2025a). Practically, this indicates that it might be worthwhile to explore interventions aimed at improving source memory in collaborative learning, as improved source memory might not impair content learning. Such interventions aimed at improving source memory could also improve content learning, as suggested by the positive correlations observed in other learning contexts. However, even in the absence of a relationship—as in the present study—better source memory may still support long-term academic help-seeking.

Regarding further variables (Need for Cognitive Closure, interest in the learning topic, and self-reported prior knowledge), learning outcome correlated positively with prior knowledge and interest in the learning topic, while there was no relationship with Need for Cognitive Closure. However, there were no relationships between source memory and further variables, even though previous studies reported positive relationships between source memory and prior knowledge (e.g., Stang Lund et al., 2019) or source memory and interest in the topic (Strømsø et al., 2010). One reason for the different results may lie in the assessment of the variables: Anmarkrud et al. (2022) have shown that relationships between sourcing and prior knowledge are more often significant when studies use multiple choice questions to assess prior knowledge rather than self-reported measures (like in the present study).

4.5 Limitations

Some limitations of this study should be acknowledged, particularly regarding the generalizability of the findings.

(1) The pseudo-collaborative nature of the study offered a highly controlled environment to isolate effects. For example, in real-collaborative contexts, group composition and learning partner expertise effects may likewise be moderated by interpersonal dynamics, such as familiarity or liking. Both learning and source memory can be affected by further aspects like phrasing or different amounts of contributions from learning partners. Our use of a pseudo-collaborative design allowed us to control these potential sources of variability by randomizing and keeping partner behavior, information distribution, and contextual features constant, thereby isolating the effects of our experimental manipulations. However, pseudo-collaborative studies also lack the interactive processes inherent to real collaboration. As discussed in earlier sections, this can also explain why the hypothesized benefits of learning in heterogeneous groups or with conflicting information were not observed in the present study. From theoretical and practical perspectives, this further suggests that the advantages of heterogeneous group compositions may rely on interactive processes such as questioning and explaining (Webb, 1989) or that benefits of conflicting information in collaborative learning may depend on mechanisms like consensus-building (Johnson and Johnson, 2009) or epistemic conflict regulation Buchs and Butera, (2004). It is also possible that the quality of social interactions may play a role in source memory formation. For example, partner modeling (i.e., mental models about learning partners' knowledge) depends on the quality of interactions between learners (Dillenbourg et al., 2016). Since partner modeling and source memory can be conceptualized as similar person-information associations from different perspectives, it can be reasonable to assume that the quality of interactions between learners could also affect how learners remember which learning partner shared certain information—an assumption to be tested in real collaborative experiments.

(2) The time between learning and testing lasted only a few minutes, while in educational practice, recalling the content or source is often required after much longer intervals—ranging from hours to days. For both learning outcome and source memory, previous findings suggest that effects of immediate tests might be generalized to longer intervals as well. For instance, acquired knowledge can become relevant days after the learning session when students take an exam. Zambrano et al. (2019) compared individual and collaborative learners across different retention intervals, including an immediate test (similar to the present experiment) and a retention test after 1 week. In both tests, learners who previously learned collaboratively outperformed those who had studied individually. This suggests that collaborative learning benefits can persist over time. Correspondingly, regarding source memory, remembering the source is sometimes required after hours or even days when learners—while rehearsing learning content—need to decide which peer to ask for more information. Nadarevic and Erdfelder (2013) have tested source memory for trivia statements after 20 min or 1 week. Although source memory declined over time, the same pattern of results emerged in both tests, i.e., a memory advantage for high- or low-credible sources compared to a medium credible source. Comparable effects have been reported for other aspects of social source memory (e.g., Buchner et al., 2009). Taken together, these findings regarding learning and source memory indicate that certain effects found in the immediate test may extend to longer retention intervals in real-world educational contexts.

Methodologically, one potential concern is that administering the source memory test prior to the knowledge test might have re-exposed participants to the content, thereby enhancing content learning. However, in our multinomial processing tree (MPT) model, all old–new recognition parameters (D) could be equated across groups, suggesting that item recognition—which could reflect content learning—was not systematically influenced by our manipulations or by the source memory test itself.

4.6 Outlook and conclusion

Note that source memory was overall relatively poor and sources were often forgotten or confused, which can have two implications. (1) Source monitoring errors can lead to the acquisition of misconceptions because learners might not double-check potential unreliable information from low-knowledge partners. Indeed, source monitoring errors can explain how unreliable information, such as fake news, can lead to false beliefs Schincariol et al., (2024). (2) Regarding academic help-seeking, learners sometimes turn to inefficient sources of help Giblin and Stefaniak, (2021), which might stem from source confusions: learners might simply forget which learning partner presented certain information and might thus ask the wrong person for help and more information. Given these potential consequences, a key question is how source memory can be supported more effectively during collaborative learning.

Previous research indicates that learners may need to be aware of their learning partners' knowledge during learning: for example, Nadarevic and Erdfelder (2013) tested source memory for social sources differing in their credibility. Participants who received information about the sources' credibility before receiving information from them remembered sources better than those who received information about their partners' credibility afterwards. Transferred to a collaborative learning context, this indicates that receiving information about partner characteristics (such as their expertise or knowledge level) before and during collaboration can improve source memory. This can be supported through cognitive group awareness tools that can provide learners with information about their partners' expertise or knowledge level (Bodemer et al., 2018). However, not all forms of knowledge visualizations may lead to the same result: Engelmann and Hesse (2010) found that concept maps displaying partners' knowledge did not improve memory for who contributed specific information. In contrast, visualizations that explicitly label the knowledge level of each learning partner—such as indicating whether someone is highly or less knowledgeable (e.g., Dehler et al., (2011); Dehler Zufferey et al., (2010)—seem to be more effective in supporting source memory, as they provide a clearer basis for distinguishing between sources. In practice, this suggests that instructors who want to strengthen students' source memory should choose group awareness tools that highlight students' knowledge levels—especially in learning subjects where conflicting information is involved, since such conflicting information can impair source memory in collaborative learning.

While accurate (source) memory can help to make adaptive social judgments in the future (Kroneisen and Bell, 2022; Schaper et al., 2019; Sklenar and Leshikar, 2025), our study did not include behavioral measures such as actual help-seeking. Thus, generalization of our findings to such learning-related strategies should be interpreted with caution. Such links could be modeled with adequate statistical power using hierarchical MPT approaches (for an overview and comparison of different MPT approaches, see Singmann et al., (2024), which can enable the estimation of individual-level memory parameters and their associations with behavioral outcomes. In the present study, our sample size and design were optimized for condition-level comparisons. However, future studies could be explicitly designed to investigate such relationships, with sufficient power and behavioral measures included.

Taken together, the present study extends the source monitoring framework (Johnson et al., 1993) to educational contexts and offers insights into source memory in collaborative learning, considering content- and person-related factors. From a methodological perspective, MPT models (which are rarely used to analyze cognitive processes in educational contexts) allowed measurements of unconfounded memory processes. Such analysis methods can also be used to assess other memory constructs crucial to collaborative learning, such as partner modeling processes (for examples, see Ülker and Bodemer, 2023, 2025b). Understanding source memory can help to derive holistic long-term implications, informing educational tool design and instructional strategies in practice. Ultimately, learners can receive support in developing sustainable collaborative practices by effectively utilizing their peers' knowledge and acquiring more (reliable) knowledge.

Data availability statement

The datasets presented in this study can be found in online repositories. The names of the repository/repositories and accession number(s) can be found in the article/Supplementary material.

Ethics statement

The studies involving humans were approved by the Local Ethics Committee of the Department of Human-Centered Computing and Cognitive Science at the University of Duisburg-Essen. The studies were conducted in accordance with the local legislation and institutional requirements. The participants provided their written informed consent to participate in this study.

Author contributions

OÜ: Writing – review & editing, Data curation, Conceptualization, Investigation, Methodology, Visualization, Writing – original draft, Formal analysis, Validation. DB: Validation, Writing – original draft, Conceptualization, Supervision, Writing – review & editing.

Funding

The author(s) declare that financial support was received for the research and/or publication of this article. Open access was supported and organized by the Open Access Funding of the University of Duisburg-Essen.

Acknowledgments

Parts of the ideas presented in this manuscript have previously appeared in a four-page short paper published in the Proceedings of the 17th International Conference on Computer-Supported Collaborative Learning (ISLS Annual Meeting): Ülker, O. and Bodemer, D. (2024). Source Memory and Collaborative Learning: The Role of Group Composition and Conflicting Information. In: eds. J. Clarke-Midura, I. Kollar, X. Gu, and C. D'Angelo. Proceedings of the 17th International Conference on Computer-Supported Collaborative Learning – CSCL 2024. International Society of the Learning Sciences. 237–240. doi: 10.22318/cscl2024.108887. The ISLS holds the copyright for the short paper. However, this manuscript in its current form has not been published before and represents a substantially extended and revised version, including different and additional analyses, more detailed descriptions of the theoretical background, methods, and results, as well as expanded interpretations and conclusions. The authors would like to thank the artists of https://www.dinosaurcouch.com/ for granting permission to use and adapt their dinosaur illustrations as avatars in this study.

Conflict of interest

The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.

Generative AI statement

The author(s) declare that Gen AI was used in the creation of this manuscript. During the preparation of this manuscript, the author(s) used (generative) AI for language improvement: DeepL for translation support, e.g., to check and optimize translations and wording, and ChatGPT (OpenAI, GPT-4) for improving clarity and optimizing phrasing. All suggestions from these tools were reviewed and revised by the author(s), who take full responsibility for the final content of the publication.

Any alternative text (alt text) provided alongside figures in this article has been generated by Frontiers with the support of artificial intelligence and reasonable efforts have been made to ensure accuracy, including review by the authors wherever possible. If you identify any issues, please contact us.

Publisher's note

All claims expressed in this article are solely those of the authors and do not necessarily represent those of their affiliated organizations, or those of the publisher, the editors and the reviewers. Any product that may be evaluated in this article, or claim that may be made by its manufacturer, is not guaranteed or endorsed by the publisher.

Supplementary material

The Supplementary Material for this article can be found online at: https://www.frontiersin.org/articles/10.3389/feduc.2025.1691038/full#supplementary-material

Footnotes

1. ^Unlike in many classical source monitoring paradigms, items were not counterbalanced as “old” or “new” across participants. As our primary aim was to simulate a realistic learning situation, we kept the set of learning texts constant to ensure comparability across all participants. “Old” items were created by extracting and paraphrasing key information units from the learning texts rather than re-presenting entire texts verbatim. This approach reflects our focus on remembering (or recognizing) the substance of information rather than its exact wording and allows us to generate more items for the source memory test. While our procedure could in principle lead to differences in old-recognition and new-recognition performance, we consider it unlikely to have introduced systematic material-based effects relevant to our main analyses, as learning texts (and thus old items in the source-memory test) were randomly assigned to the three partners for each participant.

2. ^In our study, we used a frequentist statistical framework with a complete-pooling approach that estimates memory parameters on group-level. Bayesian partial pooling approaches can estimate individual-level parameters and are particularly suitable when the main goal is to examine associations between (memory) parameters and other variables (for an overview and comparison of methods, see Singmann et al., 2024). However, these models typically require larger sample sizes and more responses per participant to yield stable estimates (Schmidt et al., 2025) and can thus be underpowered in limited sample sizes. In our experiment, we were more interested in specific memory differences (H-SM-3a and b). Here, the power to detect differences in corresponding parameters (d_High, d_Medium, d_Low) was already limited, as the corresponding trees did not have many observations, the parameters only occurred on single branches of the model, and were conditional (see also Schmidt et al., 2025). Because of our limited observations for each tree in our 2 × 2 design and the main objective to analyze specific memory differences, we opted for a frequentist framework and complete-pooling approach to maximize statistical power.

References

Anmarkrud, Ø., Bråten, I., Florit, E., and Mason, L. (2022). The role of individual differences in sourcing: a systematic review. Educ. Psychol. Rev. 34, 749–792. doi: 10.1007/s10648-021-09640-7

Crossref Full Text | Google Scholar

Bayen, U. J., Murnane, K., and Erdfelder, E. (1996). Source discrimination, item detection, and multinomial models of source monitoring. J. Exp. Psychol. Learn. Mem. Cogn. 22, 197–215. doi: 10.1037//0278-7393.22.1.197

Crossref Full Text | Google Scholar

Bell, N., Grossen, M., and Perret-Clermont, A-. N. (1985). Sociocognitive conflict and intellectual growth. New Dir. Child Adolesc. Dev. 29, 41–54. doi: 10.1002/cd.23219852905

Crossref Full Text | Google Scholar

Bell, R., Mieth, L., and Buchner, A. (2022). Coping with high advertising exposure: a source-monitoring perspective. Cogn. Res. Princ. Implic. 7:82. doi: 10.1186/s41235-022-00433-2

PubMed Abstract | Crossref Full Text | Google Scholar

Bodemer, D., Janssen, J., and Schnaubert, L. (2018). “Group awareness tools for computer-supported collaborative learning,” in International Handbook of the Learning Sciences, eds. F. Fischer, C. E. Hmelo-Silver, S. R. Goldman, and P. Reimann (Milton Park: Routledge), 351–358.

Google Scholar

Bråten, I., and Braasch, J. L. G. (2018). “The role of conflict in multiple source use,” in Handbook of Multiple Source Use, eds. I. Bråten, J. L. G. Braasch, and M. T. McCrudden (Milton Park: Routledge), 184–201.

Google Scholar

Braasch, J. L. G., Rouet, J-. F., Vibert, N., and Britt, M. A. (2012). Readers' use of source information in text comprehension. Mem. Cogn. 40, 450–465. doi: 10.3758/s13421-011-0160-6

PubMed Abstract | Crossref Full Text | Google Scholar

Braasch, J. L. G., and Scharrer, L. (2020). “The role of cognitive conflict in understanding and learning from multiple perspectives,” in Handbook of Learning from Multiple Representations and Perspectives, eds. P. Van Meter, A. List, D. Lombardi, and P. Kendeou (Milton Park: Routledge), 205–222.

Google Scholar

Bröder, A., and Meiser, T. (2007). Measuring source memory. Zeitschrift Für Psychologie/J. Psychol. 215, 52–60. doi: 10.1027/0044-3409.215.1.52

Crossref Full Text | Google Scholar

Buchner, A., Bell, R., Mehl, B., and Musch, J. (2009). No enhanced recognition memory, but better source memory for faces of cheaters. Evol. Hum. Behav. 30, 212–224. doi: 10.1016/j.evolhumbehav.2009.01.004

Crossref Full Text | Google Scholar

Buchs, C., and Butera, F. (2004). Socio-cognitive conflict and the role of student interaction in learning. New Rev. Soc. Psychol. 3, 80–87.

Google Scholar

Buder, J., and Bodemer, D. (2008). Supporting controversial CSCL discussions with augmented group awareness tools. Int. J. Comput. Support Collab. Learn. 3, 123–139. doi: 10.1007/s11412-008-9037-5

Crossref Full Text | Google Scholar

Buder, J., and Bodemer, D. (2011). “Group awareness tools for controversial CSCL discussions: dissociating rating effects and visualized feedback effects,” in Connecting Computer-Supported Collaborative Learning to Policy and Practice: CSCL2011 Conference Proceedings. Volume I—Long Papers, eds. Spada, H., Stahl, G., Miyake, N., and Law, N (Hong Kong, China: International Society of the Learning Sciences), 358–365.

Google Scholar

Crommelinck, M., and Anseel, F. (2013). Understanding and encouraging feedback-seeking behaviour: a literature review. Med. Educ. 47, 232–241. doi: 10.1111/medu.12075

PubMed Abstract | Crossref Full Text | Google Scholar

de Pereyra, G, Britt, M. A., Braasch, J. L. G., and Rouet, J. -F. (2014). Reader's memory for information sources in simple news stories: effects of text and task features. J. Cogn. Psychol. 26, 187–204. doi: 10.1080/20445911.2013.879152

Crossref Full Text | Google Scholar

Dehler Zufferey, J., Bodemer, D., Buder, J., and Hesse, F. W. (2010). Partner knowledge awareness in knowledge communication: learning by adapting to the partner. J. Exp. Educ. 79, 102–125. doi: 10.1080/00220970903292991

Crossref Full Text | Google Scholar

Dehler, J., Bodemer, D., Buder, J., and Hesse, F. W. (2011). Guiding knowledge communication in CSCL via group knowledge awareness. Comput. Human Behav. 27, 1068–1078. doi: 10.1016/j.chb.2010.05.018

Crossref Full Text | Google Scholar

Dillenbourg, P. (1999). “What do you mean by collaborative learning?” in Collaborative Learning: Cognitive and Computational Approaches, ed. P. Dillenbourg (Oxford: Elsevier), 1–19.

Google Scholar

Dillenbourg, P., Lemaignan, S., Sangin, M., Nova, N., and Molinari, G. (2016). The symmetry of partner modelling. Int. J. Comput. Support Collab. Learn. 11, 227–253. doi: 10.1007/s11412-016-9235-5

Crossref Full Text | Google Scholar

Engelmann, T., and Hesse, F. W. (2010). How digital concept maps about the collaborators' knowledge and information influence computer-supported collaborative problem solving. Int. J. Comput. Support Collab. Learn. 5, 299–319. doi: 10.1007/s11412-010-9089-1

Crossref Full Text | Google Scholar

Erdfelder, E., Auer, T-. S., Hilbig, B. E., Aßfalg, A., Moshagen, M., Nadarevic, L., et al. (2009). Multinomial processing tree models: a Review of the literature. Zeitschrift Für Psychologie/J. Psychol. 217, 108–124. doi: 10.1027/0044-3409.217.3.108

Crossref Full Text | Google Scholar

Erkens, M., Bodemer, D., and Hoppe, H. U. (2016). Improving collaborative learning in the classroom: text mining based grouping and representing. Int. J. Comput. Support. Collab. Learn. 11, 387–415. doi: 10.1007/s11412-016-9243-5

Crossref Full Text | Google Scholar

Faul, F., Erdfelder, E., Lang, A-. G., and Buchner, A. (2007). G^*Power 3: a flexible statistical power analysis program for the social, behavioral, and biomedical sciences. Behav. Res. Methods 39, 175–191. doi: 10.3758/BF03193146

PubMed Abstract | Crossref Full Text | Google Scholar

Giblin, J., and Stefaniak, J. (2021). Examining decision-making processes and heuristics in academic help-seeking and instructional environments. TechTrends 65, 101–110. doi: 10.1007/s11528-020-00556-7

Crossref Full Text | Google Scholar

Heimbuch, S., and Bodemer, D. (2017). Controversy awareness on evidence-led discussions as guidance for students in wiki-based learning. Internet High. Educ. 33, 1–14. doi: 10.1016/j.iheduc.2016.12.001

Crossref Full Text | Google Scholar

Holm, S. (1979). A simple sequentially rejective multiple test procedure. Scand. J. Stat. 6, 65–70.

Google Scholar

Janssen, J., and Bodemer, D. (2013). Coordinated computer-supported collaborative learning: awareness and awareness tools. Educ. Psychol. 48, 40–55. doi: 10.1080/00461520.2012.749153

Crossref Full Text | Google Scholar

Johnson, D. W., and Johnson, R. T. (2009). Energizing learning: the instructional power of conflict. Educ. Res. 38, 37–51. doi: 10.3102/0013189X08330540

Crossref Full Text | Google Scholar

Johnson, M. K., Hashtroudi, S., and Lindsay, D. S. (1993). Source monitoring, Psychol. Bull. 114, 3–28. doi: 10.1037/0033-2909.114.1.3

Crossref Full Text | Google Scholar

Jurica, P. J., and Shimamura, A. P. (1999). Monitoring item and source information: evidence for a negative generation effect in source memory. Mem. Cogn. 27, 648–656. doi: 10.3758/bf03211558

PubMed Abstract | Crossref Full Text | Google Scholar

Karabenick, S. A., and Berger, J-. L. (2013). “Help seeking as a self-regulated learning strategy,” in Applications of Self-Regulated Learning Across Diverse Disciplines: A Tribute to Barry J. Zimmerman, eds. H. Bembenutty, T. J. Cleary, and A. Kitsantas (Charlotte, NC: IAP Information Age Publishing), 237–261.

Google Scholar

Keefe, R. S. E., Arnold, M. C., Bayen, U. J., McEvoy, J. P., and Wilson, W. H. (2002). Source-monitoring deficits for self-generated stimuli in schizophrenia: multinomial modeling of data from three sources. Schizophr. Res. 57, 51–67. doi: 10.1016/S0920-9964(01)00306-1

PubMed Abstract | Crossref Full Text | Google Scholar

Keller, T., and Tergan, S-. O. (2005). “Visualizing knowledge and information: an introduction,” in Knowledge and Information Visualization, eds. S.-O. Tergan and T. Keller (Berlin, Heidelberg: Springer), 1–23.

Google Scholar

Kollar, I., Fischer, F., and Hesse, F. W. (2006). Collaboration scripts—a conceptual analysis. Educ. Psychol. Rev. 18, 159–185. doi: 10.1007/s10648-006-9007-2

Crossref Full Text | Google Scholar

Kroneisen, M. (2018). Is he important to me? Source memory advantage for personally relevant cheaters. Psychonom. Bull. Rev. 25, 1129–1137. doi: 10.3758/s13423-017-1345-1

PubMed Abstract | Crossref Full Text | Google Scholar

Kroneisen, M. (2024). Context matters: The influence of the social situation on source memory. Cult. Evol. 20, 111–123. doi: 10.1556/2055.2023.00039

Crossref Full Text | Google Scholar

Kroneisen, M., and Bell, R. (2022). Memory as a cognitive requirement for reciprocal cooperation. Curr. Opin. Psychol., 43, 271–277. doi: 10.1016/j.copsyc.2021.08.008

PubMed Abstract | Crossref Full Text | Google Scholar

Kuhlmann, B. G., Symeonidou, N., Tanyas, H., and Wulff, L. (2021). Remembering and reconstructing episodic context: An overview of source monitoring methods and behavioral findings. Psychol. Learn. Motiv. 75, 79–124.

Google Scholar

Maier, J., and Richter, T. (2013). Text belief consistency effects in the comprehension of multiple texts with conflicting information. Cogn. Instr. 31, 151–175. doi: 10.1080/07370008.2013.769997

Crossref Full Text | Google Scholar

Makara, K. A., and Karabenick, S. A. (2013). “Characterizing sources of academic help in the age of expanding educational technology: a new conceptual framework,” in Advances in Help-Seeking Research and Applications: The Role of Emerging Technologies, eds. S. A. Karabenick and M. Puustinen (Charlotte, NC: Information Age Publishing), 37–72.

Google Scholar

Martín-Arbós, S., Castarlenas, E., and Dueñas, J-. M. (2021). Help-seeking in an academic context: a systematic review. Sustainability, 13, 4460. doi: 10.3390/su13084460

Crossref Full Text | Google Scholar

McGrew, S., Reynolds, E. C., and Glass, A. C. (2024). The problem with perspective: Students' and teachers' reasoning about credibility during discussions of online sources. Cogn. Instr. 42, 399–425. doi: 10.1080/07370008.2024.2340981

Crossref Full Text | Google Scholar

Mieth, L., Schaper, M. L., Kuhlmann, B. G., and Bell, R. (2021). Memory and metamemory for social interactions: Evidence for a metamemory expectancy illusion. Mem. Cogn. 49, 14–31. doi: 10.3758/s13421-020-01071-z

PubMed Abstract | Crossref Full Text | Google Scholar

Moshagen, M. (2010). multiTree: a computer program for the analysis of multinomial processing tree models. Behav. Res. Methods 42, 42–54. doi: 10.3758/BRM.42.1.42

PubMed Abstract | Crossref Full Text | Google Scholar

Mugny, G., and Doise, W. (1978). Socio-cognitive conflict and structure of individual and collective performances. Eur. J. Soc. Psychol. 8, 181–192. doi: 10.1002/ejsp.2420080204

Crossref Full Text | Google Scholar

Murnane, K., and Bayen, U. J. (1996). An evaluation of empirical measures of source identification. Mem. Cogn. 24, 417–428. doi: 10.3758/BF03200931

PubMed Abstract | Crossref Full Text | Google Scholar

Nadarevic, L., and Erdfelder, E. (2013). Spinoza's error: Memory for truth and falsity. Mem. Cogn. 41, 176–186. doi: 10.3758/s13421-012-0251-z

PubMed Abstract | Crossref Full Text | Google Scholar

Nadarevic, L., and Erdfelder, E. (2019). More evidence against the Spinozan model: cognitive load diminishes memory for “true” feedback. Mem. Cogn. 47, 1386–1400. doi: 10.3758/s13421-019-00940-6

PubMed Abstract | Crossref Full Text | Google Scholar

Newman, R. S. (2002). How self-regulated learners cope with academic difficulty: the role of adaptive help seeking. Theory Pract. 41, 132–138. doi: 10.1207/s15430421tip4102_10

Crossref Full Text | Google Scholar

Odo, C., Masthoff, J., and Beacham, N. (2019). “Group formation for collaborative learning: a systematic literature review,” in Artificial Intelligence in Education, eds. S. Isotani, E. Millán, A. Ogan, P. Hastings, B. McLaren, and R. Luckin (Berlin: Springer International Publishing), 206–212.

Google Scholar

Old, S. R., and Naveh-Benjamin, M. (2008). Differential effects of age on item and associative measures of memory: a meta-analysis. Psychol. Aging 23, 104–118. doi: 10.1037/0882-7974.23.1.104

PubMed Abstract | Crossref Full Text | Google Scholar

Oshima, J., Ohsaki, A., Yamada, Y., and Oshima, R. (2017). “Collective knowledge advancement and conceptual understanding of complex scientifc concepts in the jigsaw instruction” in Making a diference: Prioritizing equity and access in CSCL, 12th International Conference on Computer Supported Collaborative Learning (CSCL) 2017, eds. B. K. Smith, M. Borge, E. Mercier, and K. Y. Lim (International Society of the Learning Sciences), 57–64.

Google Scholar

Palinscar, A. S., and Brown, A. L. (1984). Reciprocal teaching of comprehension-fostering and comprehension-monitoring activities. Cogn. Instr. 1, 117–175. doi: 10.1207/s1532690xci0102_1

Crossref Full Text | Google Scholar

Pena, M. M., Klemfuss, J. Z., Loftus, E. F., and Mindthoff, A. (2017). The effects of exposure to differing amounts of misinformation and source credibility perception on source monitoring and memory accuracy. Psychol. Conscious. Theory Res. Pract. 4, 337–347. doi: 10.1037/cns0000137

Crossref Full Text | Google Scholar

Pepe, N. W., Wang, Q., and Rajaram, S. (2021). Collaborative remembering in ethnically uniform and diverse group settings. J. Appl. Res. Mem. Cogn. 10, 95–103. doi: 10.1016/j.jarmac.2020.08.001

Crossref Full Text | Google Scholar

Perfetti, C. A., Rouet, J.-F., and Britt, M. A. (1999). “Toward a theory of documents representation,” in The Construction of Mental Representations During Reading, eds. H. van Oostendorp and S. R. Goldman (Mahwah, NJ: Lawrence Erlbaum Associates), 99–122.

Google Scholar

Perret-Clermont, A-. N. (2022). “Socio-cognitive conflict,” in The Palgrave Encyclopedia of the Possible, eds. V. P. Glăveanu (London: Palgrave Macmillan), 1–8.

Google Scholar

Ploetzner, R., Dillenbourg, P., Preier, M., and Traum, D. (1999). Learning by explaining to oneself and to others. Collab. Learn. Cogn. Comput. Approach. 1, 103–121.

Google Scholar

Pornpitakpan, C. (2004). The persuasiveness of source credibility: a critical review of five decades' evidence. J. Appl. Soc. Psychol. 34, 243–281. doi: 10.1111/j.1559-1816.2004.tb02547.x

Crossref Full Text | Google Scholar

Razmerita, L., and Brun, A. (2011). “Collaborative learning in heterogeneous classes: towards a group formation methodology,” in Proceedings of the 3rd International Conference on Computer Supported Education, eds. A. Verbraeck, M. Helfert, J. Cordeiro, and B. Shishkov (Setúbal: SciTePress), 189–194.

Google Scholar

Sangin, M., Molinari, G., Nüssli, M.-A., and Dillenbourg, P. (2011). Facilitating peer knowledge modeling: Effects of a knowledge awareness tool on collaborative learning outcomes and processes. Comput. Hum. Behav. 27, 1059–1067. doi: 10.1016/j.chb.2010.05.032

Crossref Full Text | Google Scholar

Schacter, D. L. (1999). The seven sins of memory: insights from psychology and cognitive neuroscience. Am. Psychol. 54, 182–203. doi: 10.1037/0003-066X.54.3.182

PubMed Abstract | Crossref Full Text | Google Scholar

Schaper, M. L., Mieth, L., and Bell, R. (2019). Adaptive memory: source memory is positively associated with adaptive social decision making. Cognition 186, 7–14. doi: 10.1016/j.cognition.2019.01.014

PubMed Abstract | Crossref Full Text | Google Scholar

Schincariol, A., Otgaar, H., Greene, C. M., Murphy, G., Riesthuis, P., Mangiulli, I., et al. (2024). Fake memories: a meta-analysis on the effect of fake news on the creation of false memories and false beliefs. Mem. Mind Media 3:e17. doi: 10.1017/mem.2024.14

Crossref Full Text | Google Scholar

Schlink, S., and Walther, E. (2007). Kurz und gut: Eine deutsche Kurzskala zur Erfassung des Bedürfnisses nach kognitiver Geschlossenheit. Zeitschrift für Sozialpsychologie 38, 153–161. doi: 10.1024/0044-3514.38.3.153

Crossref Full Text | Google Scholar

Schlusche, C., Schnaubert, L., and Bodemer, D. (2021). Perceived social resources affect help-seeking and academic outcomes in the initial phase of undergraduate studies. Front. Educ. 6:732587. doi: 10.3389/feduc.2021.732587

Crossref Full Text | Google Scholar

Schlusche, C., Schnaubert, L., and Bodemer, D. (2024). Competence in (meta-)cognitive learning strategies during help-seeking to overcome knowledge-related difficulties. Active Learn. High. Educ. 25, 489–515. doi: 10.1177/14697874231168343

Crossref Full Text | Google Scholar

Schmidt, O., Erdfelder, E., and Heck, D. W. (2025). How to develop, test, and extend multinomial processing tree models: a tutorial. Psychol. Methods 30, 720–743. doi: 10.1037/met0000561

PubMed Abstract | Crossref Full Text | Google Scholar

Schnaubert, L., and Bodemer, D. (2022). Group awareness and regulation in computer-supported collaborative learning. Int. J. Comput. Support. Collab. Learn. 17, 11–38. doi: 10.1007/s11412-022-09361-1

Crossref Full Text | Google Scholar

Singmann, H., Heck, D. W., Barth, M., Erdfelder, E., Arnold, N. R., Aust, F., et al. (2024). Evaluating the robustness of parameter estimates in cognitive models: a meta-analytic review of multinomial processing tree models across the multiverse of estimation methods. Psychol. Bull. 150, 965–1003. doi: 10.1037/bul0000434

PubMed Abstract | Crossref Full Text | Google Scholar

Sklenar, A. M., and Leshikar, E. D. (2025). Memory as a foundation for approach and avoidance decisions: a fertile area for research. Mem. Cogn. 53, 590–605. doi: 10.3758/s13421-024-01588-7

PubMed Abstract | Crossref Full Text | Google Scholar

Snodgrass, J. G., and Corwin, J. (1988). Pragmatics of measuring recognition memory: applications to dementia and amnesia. J. Exp. Psychol. Gener. 117, 34–50. doi: 10.1037/0096-3445.117.1.34

PubMed Abstract | Crossref Full Text | Google Scholar

Soller, A. (2001). Supporting social interaction in an intelligent collaborative learning system. Int. J. Artif. Intell. Educ. 12, 40–62.

Google Scholar

Stang Lund, E., Bråten, I., Brandmo, C., Brante, E. W., and Strømsø, H. I. (2019). Direct and indirect effects of textual and individual factors on source-content integration when reading about a socio-scientific issue. Read. Writ. 32, 335–356. doi: 10.1007/s11145-018-9868-z

Crossref Full Text | Google Scholar

Strømsø, H. I., Bråten, I., and Britt, M. A. (2010). Reading multiple texts about climate change: the relationship between memory for sources and text comprehension. Learn Instr. 20, 192–204. doi: 10.1016/j.learninstruc.2009.02.001

Crossref Full Text | Google Scholar

Strømsø, H. I., Bråten, I., Britt, M. A., and Ferguson, L. E. (2013). Spontaneous sourcing among students reading multiple documents. Cogn. Instr. 31, 176–203. doi: 10.1080/07370008.2013.769994

Crossref Full Text | Google Scholar

Symeonidou, N., and Kuhlmann, B. G. (2021). A novel paradigm to assess storage of sources in memory: the source recognition test with reinstatement. Memory 29, 507–523. doi: 10.1080/09658211.2021.1910310

PubMed Abstract | Crossref Full Text | Google Scholar

Thomm, E., and Bromme, R. (2016). How source information shapes lay interpretations of science conflicts: interplay between sourcing, conflict explanation, source evaluation, and claim evaluation. Read. Writ. 29, 1629–1652. doi: 10.1007/s11145-016-9638-8

Crossref Full Text | Google Scholar

Ülker, O., and Bodemer, D. (2023). “Remembering the knowledge of experts and novices in computer-supported collaborative learning: a multinomial processing tree approach,” in Proceedings of the 31st International Conference on Computers in Education, eds. JL. Shih, A. Kashihara, W. Chen and H. Ogata (Taoyuan: Asia-Pacific Society for Computers in Education), 286–291.

Google Scholar

Ülker, O., and Bodemer, D. (2025a). “'Memorize the source!'—Does focusing on learning partners improve source memory without impairing content learning?,” in Proceedings of the 18th International Conference on Computer-Supported Collaborative Learning—CSCL 2025, eds. Oshima, J., Chen, B., Vogel, F., and Järvelä, J (Hong Kong: International Society of the Learning Sciences), 90–98.

Google Scholar

Ülker, O., and Bodemer, D. (2025b). Selecting learning partners: memory for participation and competence. Cogn. Res. Princ. Implic. 10:46. doi: 10.1186/s41235-025-00656-z

PubMed Abstract | Crossref Full Text | Google Scholar

Vogel, F., Wecker, C., Kollar, I., and Fischer, F. (2017). Socio-cognitive scaffolding with computer-supported collaboration scripts: a meta-analysis. Educ. Psychol. Rev. 29, 477–511. doi: 10.1007/s10648-016-9361-7

Crossref Full Text | Google Scholar

Webb, N. M. (1989). Peer interaction and learning in small groups. Int. J. Educ. Res. 13, 21–39. doi: 10.1016/0883-0355(89)90014-1

Crossref Full Text | Google Scholar

Webster, D. M., and Kruglanski, A. W. (1994). Individual differences in need for cognitive closure. J. Pers. Soc. Psychol. 67, 1049–1062. doi: 10.1037/0022-3514.67.6.1049

PubMed Abstract | Crossref Full Text | Google Scholar

Wertgen, A. G., and Richter, T. (2023). Source credibility and plausibility are considered in the validation of textual information: evidence from a social media context. J. Cogn. Psychol. 35, 183–204. doi: 10.1080/20445911.2022.2149757

Crossref Full Text | Google Scholar

Zamani, M. (2016). Cooperative learning: homogeneous and heterogeneous grouping of Iranian EFL learners in a writing context. Cogent. Educ. 3:1149959. doi: 10.1080/2331186X.2016.1149959

Crossref Full Text | Google Scholar

Zambrano, R., Kirschner, J., Sweller, F. J., and Kirschner, P. A. (2019). Effects of prior knowledge on collaborative and individual learning. Learn Instr. 63:101214. doi: 10.1016/j.learninstruc.2019.05.011

Crossref Full Text | Google Scholar

Zhang, L., Kalyuga, S., Lee, C. H., Lei, C., and Jiao, J. (2015). “Effectiveness of collaborative learning with complex tasks under different learning group formations: a cognitive load perspective,” in Hybrid Learning: Innovation in Educational Practices, eds. S. K. S. Cheung, L. Kwok, H. Yang, J. Fong, and R. Kwan (Berlin: Springer International Publishing), 149–159.

Google Scholar

Appendix A

Memory measures

(1) Conditional Source Identification Measures (CSIM, Bröder and Meiser, 2007; Murnane and Bayen, 1996) are calculated by dividing the number of correct source judgments (see Table A1) by the number of hits (i.e., the corresponding sum of “high”, “medium”, and “low” or “A”, “B”, and “C” responses). For example, to calculate the identification measure for the high-knowledge partner as a source of information in the “heterogeneous group and with conflict” condition, one would use the formula:

$\begin{array}{l}\frac{{\text{High}}_{\text{High}}}{{\text{High}}_{\text{High}}+{\text{Medium}}_{\text{High}}+{\text{Low}}_{\text{High}}}=\frac{155}{155+106+74}=0.46\end{array}$

Where:

• High_High = Number of “high” responses for information originally presented by the high-knowledge partner (correct responses)

• Medium_High = Number of “medium” responses for information originally presented by the high-knowledge partner (incorrect responses)

• Low_High = Number of “low” responses for information originally presented by the high-knowledge partner (incorrect responses)

To give another example, the identification measure for partner C in the “homogeneous group and without conflict” condition would be calculated by:

$\begin{array}{l}\frac{{\text{C}}_{\text{C}}}{{\text{A}}_{\text{C}}+{\text{B}}_{\text{C}}+{\text{C}}_{\text{C}}}=\frac{133}{94+91+133}=0.41\end{array}$

Where:

• A_C = Number of “A” responses for information originally presented by partner C (incorrect responses)

• B_C = Number of “B” responses for information originally presented by partner C (incorrect responses)

• C_C = Number of “C” responses for information originally presented by partner C (correct responses)

Table A1

Table A1. Aggregated response frequencies in the source memory test phase.

Mean CSIM scores are depicted in Figure 5. While our examples illustrate the calculation on a group level, CSIM scores are calculated on an individual level and such classification-based scores can be obtained for each participant, which therefore allows testing for statistical relationships between source identification (as a proxy for source memory) and other measures (such as learning outcome or prior knowledge). However, such classification-based measures can also confound memory and guessing processes (Bröder and Meiser, 2007) and should be interpreted with caution.

(2) Multinomial processing tree models (for an overview, see Erdfelder et al., 2009) offer a solution and disentangle memory and guessing processes. We used the two-high threshold model of source monitoring (Murnane and Bayen, 1996), adapted for our study purposes. Also, we used a frequentist statistical framework (for an overview, see Singmann et al., 2024) and a complete pooling approach to increase statistical power (see Schmidt et al., 2025). Our model contains 12 parameters per condition (D_A/High, D_B/Medium, D_C/Low, D_N, d_A/High, d_B/Medium, d_C/Low, a_A/High, a_B/Medium, g_A/High, g_B/Medium, b), which represent probabilities of certain cognitive processes, ranging from 0 to 1. To analyze the effects of conflicting information and group composition on source memory, we used four sets of the model depicted in Figure 4, with one set for each of the four experimental conditions. Each processing tree represents one class of information (original source A/High, B/Medium, C/Low, new information).

To illustrate, in either condition in a homogeneous group with partners with same knowledge levels, the first tree of the model depicts a situation in the test phase in which participants have to judge the source of a piece of information which was originally presented by partner A in the learning phase. Several cognitive processes (along the lines) lead to the different possible classifications the participant can make (rectangles on the right side). With the probability D_A, the information will be detected as old. If the information is detected as old, participants can recognize correctly that the information was presented by partner A with the conditional probability d_A, which therefore reflects source memory. With the complementary probability 1 – d_A, participants do not remember the source and have to rely on guessing processes: participants might guess the incorrect source “partner B” with the probability a_B. Given partner B was not guessed (1 – a_B), participants can guess the correct source (partner A) with the probability a_A or another incorrect source (partner C) with the probability (1 – a_A). If a piece of information is not detected as old (with the probability 1 – D_A), it might be correctly guessed as old with the probability b. Following an “old” guess, participants also have to guess the source of the information, which is analogue to the previous described guessing processes, but represented by the letter g instead of a. Alternatively, participants might incorrectly guess that the information was not presented during the learning phase with the probability 1 – b and assume the information to be new. For example, the following equation depicts the probability of giving the answer “partner A” in the source memory test to an item that was originally presented by partner A: P(“Partner A”|Partner A) = D_A × d_A + D_A × (1 – d_A) × (1 – a_B) × a_A + (1 – D_A) × b × (1 – g_B) × g_A. The bottom tree refers to information which had not been presented during the learning phase. Here, D_N is the probability of detecting that a distractor item is new. This example illustrates that (combinations of) several cognitive processes can result in the same response categories, and consequently, that a correct answer in the source memory test does not necessarily mean that the source was remembered.

Comparing empirically observed category frequencies with predicted category frequencies, one can test whether the model fits the data. Hypotheses are tested by imposing additional parameter restrictions on the base model. Whether restrictions fit the data is expressed in ΔG² (e.g., a significant increase in model misfit indicates that additional restrictions are not appropriate). While only source memory parameters (d) and some guessing parameters are analyzed, for the sake of completeness, all parameter estimates are depicted in Table A2.

Table A2

Table A2. Parameter estimates for old–new recognition, source memory, and guessing biases as a function of group composition (heterogeneous vs. homogeneous) and conflicting information (with conflict vs. without conflict).