On the social nature of abstraction: cognitive implications of interaction and diversity

Abstract

The human capacity for abstraction is remarkable. We effortlessly form abstract representations from varied experiences, generalizing and flexibly transferring experiences and knowledge between contexts, which can facilitate reasoning, problem solving and learning across many domains. The cognitive process of abstraction, however, is often portrayed and investigated as an individual process. This paper addresses how cognitive processes of abstraction—together with other aspects of human reasoning and problem solving—are fundamentally shaped and modulated by online social interaction. Starting from a general distinction between convergent thinking, divergent thinking and processes of abstraction, we address how social interaction shapes information processing differently depending on cognitive demands, social coordination and task ecologies. In particular, we suggest that processes of abstraction are facilitated by the interactive sharing and integration of varied individual experiences. To this end, we also discuss how the dynamics of group interactions vary as a function of group composition; that is, in terms of the similarity and diversity between the group members. We conclude by outlining the role of cognitive diversity in interactive processes and consider the importance of group diversity in processes of abstraction.

This article is part of the theme issue ‘Concepts in interaction: social engagement and inner experiences’.

1. Introduction

Abstraction is fundamental to many aspects of human cognition, allowing our cognitive system to group varied experiences as tokens of the same abstract type [1–3]. This makes our categories resilient to noise and enables us to respond to novel experiences in an uncertain or changing environment due to their higher order similarities or analogies (i.e. their family resemblances) to known entities [4]. Abstraction processes have been documented across numerous domains and modalities [5], and are often portrayed as generalizations across accumulated experiences in the mind of an individual [6–10], allowing representations to transfer and apply more flexibly to new or changing contexts [2,3,11,12].

In this paper, we claim that a similar, but socially distributed cognitive process can unfold in contexts of dialogical social interaction [13,14]. Here, abstract representations are the product of integration and generalization across experiences of different individuals, negotiated through social information exchange, and constituted by the summed variance of individual perspectives. In other words, we argue that social interaction stimulates cognitive processes of abstraction at the level of the group.

In stating our case, we propose a broad conceptual distinction of types of cognitive processes afforded by interactive tasks between convergent thinking, divergent thinking, and abstraction. In doing so, we aim, in part, to convey how processes of abstraction are distinctively used, in light of its counterparts.

An important corollary of the central claim of this paper stems from the finding that cognitive diversity and alignment may be important predictors of group performance in joint problem solving. In particular, interactive processes of abstraction can depend on variability in how problems are mentally represented and the corresponding dialogical contributions [14–19]. This observation is in line with evidence from group interactions more broadly, suggesting that the composition of members in a group, particularly in terms of diversity of their members, often plays a critical role [20]. To address this aspect, we conclude by considering the dynamics of group interactions which are sensitive to group composition, primarily in terms of alignment, information sharing and communication of reliability.

Thus, two of the challenges in addressing the social nature of abstraction are, on the one hand, the variability in tasks that interacting individuals face, affording different coordination demands and types of information processing. And on the other hand, the inherent variability within members of social groups, influencing the collaboration on several levels. We recognize the complexity of group cognition and the many emergent properties of the interaction themselves [21,22], and the aforementioned challenges put forward are by no means intended to be exhaustive. Rather, our account aims to inform and motivate investigations of the nature of abstraction in interaction to not only differentiate types of cognitive processing that are afforded by interactions, but also group composition and cognitive diversity.

Notice also that the term abstraction here is used as distinct from abstractness [23]. While the former describes the processes leading to particular properties of categories (e.g. the way the category DOG comes to refer to a large set of referents, from chihuahua to grand denoise, that radically vary in their perceptual properties), the latter relates to properties of particular classes of linguistic concepts such as ‘freedom’, ‘responsibility’ or ‘democracy’ [24]. It is important to note, however, that the two are inherently inter-connected [25], and that the critical role of social interaction in the acquisition and grounding of abstract concepts has been addressed in detail elsewhere [23]. Similarly, a wider link between human sociality and capacities for abstract thinking has implications for the evolution of human reasoning and language on broader cultural time scales [26–28]. In this paper, however, we focus on aspects of online cognition and behaviour in contexts of situated problem-solving and decision-making tasks. Our scope is thus limited to empirical studies of task-oriented social interactions, which allow us to spotlight cognitive processes in dynamic interactions and, essentially, how human thinking is altered as a result of these interactions.

We begin by introducing basic concepts and defining features in regard to social interactions, which inform our subsequent conceptualization of interactive dynamics in the empirical literature.

2. Conceptual features of social interactions

(a) . Interaction alters mental representations

When individuals interact to accomplish a shared goal, they automatically engage in some degree of alignment of their mental representations [29–31]. Indeed, in cognitive science, human sociality has often been proposed to be explained with reference to our unique capacities for inferring and sharing the mental states of others or engaging in some form of shared intentionality [32–34]. One of the key questions for research in this domain has been how exactly interaction alters mental representation in individuals, that is, how we interact and adapt to each other to form shared representations of the world [35].

Elaborate theoretical answers to this key question is beyond the scope of the present paper, but it is relevant to note that a prominent approach in the literature has been to operationalize the degree of reciprocity and presence of shared goals as the main defining features of social interaction [36–39], which enable individuals to form shared representations, obtain greater understanding of others, and jointly gain access to more representational information [40,41].

(b) . Shared representations and information exchange

An influential approach to the problem of shared representation is the theoretical concept of the we-mode; an irreducibly collective perspective or state that individuals may enter into automatically when engaging in a joint action or task [42]. Here, a given action is represented as pursued together with others and entails that individuals gain access to additional information from other group members. More recently, however, and important in regard to the central argument of the present paper, this approach has taken a step away from the notion of shared goals as the paradigm case for shared representations. Instead, the focus is shifted onto the nature of the information exchange itself between interacting individuals, achieved by adjusting and aligning mental states to varying degrees [35]. Such mutual adaptations, or reciprocal forms of alignment, can potentially be conducive to a high amount of information to be exchanged. While avoiding making assumptions about the representational configuration of individual mental states, the emphasis is rather on the degree of shared information available and the manner in which shared information is socially processed, integrated and implemented, irrespective of whether there are shared goals or not.

In other words, following this approach we are particularly interested in the degree of mutual adaptation and bidirectional information exchange, which implies treating shared information processing as a continuum (rather than a discrete all-or-none phenomenon) [35]. Importantly, this shifts the focus from the nature of individual-level representations, the goal-oriented nature of a given task, or the specific representational configuration of intentional states, to the dynamical context of how individuals construe and process shared information. In this view, we integrate the nature of the interaction and the corresponding cognitive processes that are stimulated and afforded [35].

By ‘interactive tasks’, we refer to online, dialogical activities, in which individuals coordinate or collaborate and are mutually and explicitly aware of the task. To present our argument concisely, we limit these cases to involve, for instance, judgement and decision-making, problem-solving, categorization, brainstorming and creative ideation tasks, etc. When addressing these types of cognitive activities, it can be useful to evoke the metaphor of a solution space (also sometimes referred to as an action space), in which the possible solutions to a task are mapped out, and which is searched by the individuals as they navigate the shared information of the task [43–45]. When we occasionally fall back on the term representation, we use it broadly to refer to solution spaces, which can be mentally construed and explored for an ideal or multiple solutions. That is, when individuals are said to ‘represent’ task-relevant information, it is intended in this functional sense.

From this conceptual outset, we will turn to the empirical and more concrete instantiation of these notions. We will consider how information processing unfolding at the level of the group is shaped by the nature of the task and aspects of the interaction.

3. Types of information processing in interactive tasks

In conceptual models of interactive behaviour, it is often implicitly assumed that the mode of information processing is fundamentally similar across various tasks or coordination demands: individuals align their mental states, coordinate or mutually adapt to each other with varying degree of success and behavioural outcomes [30,46]. However, we suggest that a relevant distinction can be made between the types of information processing that are afforded by the particular task at hand. That is, the processes that are inherently afforded by different coordination problems can be grouped at a mechanistic level, according to the ecology of the task and its particular cognitive demands.

Here, we make a first attempt at discerning the way that social interaction stimulates processes specific to different broad types of cognitive problems. While a more fine-grained typology of cognitive processing can likely be used, we distinguish between divergent and convergent thinking [47,48]. In addition, we consider in more detail processes of abstraction. Although the latter might not constitute a type that is fully perpendicular to divergent and convergent thinking, we emphasize conditions where cognitive processes of abstraction, in particular, are afforded, which warrants a separate conceptual treatment.

We define convergent thinking as processes that are inherently afforded by problems or tasks that require participants to delimit the possible set of solutions down to one particular domain or a single ideal solution. In contrast, processes of divergent thinking are afforded by problems that require participants to flexibly expand or multiply possible solutions. Processes of abstraction, on the other hand, are inherently afforded by problems that require participants to uncover an underlying organizing parameter, or generalize and transfer a possible solution between contexts. In the latter case, we recognize that processes of abstraction may occur in tasks that involve convergent or divergent thinking. However, we argue that the cognitive processes involved in abstraction stand out in a number of ways and the distinctions put forward here relate primarily to types of information processing more than types of concrete tasks. We will briefly elaborate on these processes in the following.

(a) . Convergent thinking

Interaction can stimulate processes of convergent thinking when problems require a single objectively true answer, a fixed solution or a hierarchy of increasingly optimal solutions. Convergent thinking is thus typically afforded when interacting individuals are attempting to socially align on decision-making, search or categorization tasks, insight problems, etc. Taken together, these interactive situations are characterized by individuals jointly exploring a mental solution space, in a way that reduces uncertainty and continuously narrows down the field of competing candidate solutions through learning and optimization.

The possible benefits of group interaction are a familiar but also contentious concept, not without complications. Condorcet's Jury Theorem was one of the first attempts to mathematically formalize the idea that communication with others (i.e. comparing and contrasting experiences) could improve understanding and convergence on more ideal decisions and judgements [49]. On the other hand, the first empirical evidence that combining information across independent individuals leads to more reliable information was demonstrated by Galton [50]. However, this type of pooling of information assumes that individuals' contributions are entirely independent—an assumption that can rarely be relied on—and that precludes interaction by definition. These early efforts have, more recently, led to a large body of research devoted to elaboration of the wisdom-of-crowds effect [51,52], group decision-making [53], social problem solving [13] and convergence dynamics [54,55], and to develop the recipe for how exactly to converge and combine social information in order to optimize toward a consensus decision or strategy [56].

For example, to converge in group decision-making, a key ingredient is the process of evaluating and weighting the competence and reliability of others’ information [57–59]. Given that individuals usually do not have direct or objective access to the reliability of others, estimating reliability is often based on more indirect social markers pertaining to perceived expertise [60], status [61] or confidence [62]. Importantly, such markers themselves can sometimes be misleading, expressing a level of reliability which does not correspond to an individual's actual ability [63,64]. Indeed, in group interactions, a recurrent challenge consists of accommodating individual variability in reliability, as there can be hidden biases or heuristic strategies at play [65,66].

Other key ingredients in tasks involving processes of convergent thinking include how individual group members represent problems [67,68], and their level of metacognitive awareness [69]. For instance, for a broader cognitive search to converge on an ideal solution, groups might benefit from the right combination of members' cognitive styles: some group members being ‘explorers’ (searching for new regimes of the solution space), while others are ‘exploiters’ (carefully assessing candidate solutions already accessible in the current regime) [67]. In addition, individual group members' ability to explicitly estimate the accuracy of their own information or performance (i.e. metacognitive bias) is important for the group's calibration of certainty and reliability [70]. The relative importance of such processes in the interaction depends on the composition of the group and relations between members of the group, which we will return to subsequently.

A particularly successful format of joint convergent thinking seems to be when individuals through repeated interactions openly discuss and integrate their experiences and knowledge in order to exchange and integrate information [13,22,71–73]. This research suggests that repeated interactions, in which members of a group attempt to reach consensus, allow for the assessment of the reasoning behind, or underlying basis of individual and shared knowledge, and for the mutual understanding to be made explicit. This alleviates social influence (referred to as bias) from ‘seemingly’ competent information and misunderstandings. As a result, more weight is put on the well-reasoned and unbiased information rather than relying on indirect social markers of reliability. For example, the collective accuracy of a large crowd at a TEDx event (answering general knowledge questions) was increased considerably by simply dividing the crowd into smaller groups that shared arguments and reasoned together [74]. Here, the group consensus answers, on average, were less biased and resulted in decreased variance compared to initial individual answers. The study suggests that biases can be mitigated by making them explicit through dialogical interaction, perhaps because biases are more obvious to others [75,76], which avoids giving undue weight to less reliable information [56].

In other words, through iterative interactions, individuals gain access not only to social information but also how that information was obtained or how an individuals' understanding was reached, potentially revealing poorly based thinking. Thus, markers of reliability which may have been misleading, such as high-status (or highly confident) individuals with poor competence, will then be subject to a ‘levelling out’ process. This recalibration of reliability increases the likelihood of the combined information to be correct. Interestingly, the benefit of convergence can in some cases also carry over to subsequent individual contexts [74,77].

(b) . Divergent thinking

Divergent thinking is used about tasks that require individuals or groups to come up with multiple solutions to a prompt. These include contexts of brainstorming, verbal fluency or open-ended creative ideation tasks, for instance, in contexts of creative activities or design. In these cases, performance is often contingent on the ability to flexibly engage or shift between multiple perspectives, strategies or methods to overcome cognitive fixation and broaden exploratory search. In empirical approaches, divergent thinking is often operationalized in three complementary measures, fluency (the number of candidate solutions produced by a participant), flexibility (the difference between solutions, indicative of the participant's ability to broadly explore a solution space) and originality (the relative frequency or ‘rareness’ of suggested solutions). Utilizing the metaphor of search in a ‘space of possible solutions’, some solutions may appear ‘in close proximity’, meaning that they readily and immediately come to mind, while others are ‘distant’, ‘hidden’ or ‘hard to find’. The mental search for solutions can thus be described (and even formalized) with analogy to animal foraging behaviours [78].

Foraging processes often present particular dynamical patterns consisting of successions of short and long paths through an area of scattered resources [45,79]. Here, a short path connects closely related solutions belonging to the same domain, category or kind, while a long path signifies a movement to a solution that is relatively distant from the last solution. The divergent thinking process thus unfolds in a succession of phases consisting of exploitation phases, visiting multiple closely related solutions, and exploration phases, consisting of longer paths to relatively unrelated solutions [44]. The relative balance between exploitation and exploration has been found to affect task performance in divergent thinking tasks [78,80], While high degrees of exploitation seem to positively affect the relative fluency of divergent thinking, flexibility and originality seems to be associated with the propensity and ability to also explore distal corners of the solutions space [81].

Social interaction has been found to affect properties and outcomes of divergent thinking processes. However, studies present mixed findings depending, for instance, on the operationalization of variables and measures. When assessing the relative fluency (number of candidate solutions) of a divergent thinking task, we can either compare the group to the performance of the best group member when she performs the task alone. But we can also make a comparison to the so-called ‘nominal group’, that is, a post hoc concatenation of unique solutions provided by all group members when they work alone [80]. In a series of studies of brainstorming sessions, it has been found that groups will often have a collective benefit compared to the best group member, but will be outperformed by the nominal group, pointing to the potentially inhibitory effects of social interactions in these contexts [82–85]. While social interaction can assist group members escape cognitive fixedness and broaden their search, it can also be potentially disruptive to individual search processes and thus prevent the group from fulfilling its full potential [83]. However, while these observations primarily address the fluency of divergent thinking processes, it is less clear how interaction affects aspects of flexibility and originality.

We speculate that social interaction stimulates cognitive flexibility and thereby affects the balance between exploitation and exploration. The individual divergent thinker might thus—in the outset—be biased toward exploitation (subject to individual differences [86]), making them more prone to cognitive fixedness: the propensity to get stuck in a local minimum and unable to identify a more distal cluster of solutions. Dialogical interaction seems to facilitate exactly this. When group members share their intuitions, perspectives and strategies, they are more likely to escape cognitive fixation, shifting the balance toward a more exploratory search style (more ‘long jumps’ in the solution space). However, this can have the disruptive effect that the group leaves clusters of potential solutions prematurely, as new ideas or intuitions continuously bring the group to different areas of the solutions space. A search style biased towards exploration thus presents a more ‘jumpy’ search pattern, which might result in better overall coverage of the solution space, but does not necessarily benefit fluency.

A number of recent studies seem to support these ideas. In an agent-based simulation, Rocca & Tylén [87] compared individual and pairs of agents as they conducted a verbal fluency task: to name as many animals as possible before experiencing cognitive fixation. The findings suggest that social interaction indeed changes the internal dynamics of divergent thinking. Contingent on the degree of agents' cognitive diversity, pairs tended to exhibit a more exploratory search style, characterized by longer jumps in semantic space. With moderate levels of diversity, pairs outperformed individuals both in terms of fluency, flexibility and originality (the relative frequency of animals relative to all animals named in the simulation). However, with high levels of diversity, the search became too ‘jumpy’, causing performance to break down.

The intimate relation between social interaction and unfolding processes of divergent thinking is also investigated in Rosenberg et al. [81]. Pairs of participants were instructed to jointly build ‘interesting and beautiful’ figures by moving around a set of tiles on an 80-inch touchscreen and saving them to their ‘gallery’ [79,88]. The experimental paradigm allowed the authors to quantify aspects of the social coordination between group members and relate these to their search behaviours. The study found that the style of interaction had a systematic impact on the dynamics and results of cognitive search. Groups characterized by dominance (one group member conducts most of the tile moves), were characterized by a more exploitative search style and higher fluency (more resulting figures), while groups characterized by a turn-taking dynamics (taking turns to move the tiles), displayed a more exploratory search behaviour and produced more original figures.

(c) . Interaction and the process of abstraction

Another way that social interaction affects cognition is by stimulating cognitive processes of abstraction. Capacities for abstract reasoning underlie many aspects of human innovation and problem solving. Processes of abstraction entails the recognition and profiling of (hidden) central organizing dimensions of a phenomenon, while backgrounding or ignoring coincidental surface details. Abstraction thus allows us to generalize and flexibly transfer experiences and knowledge between contexts by mapping their higher-level analogies, which in turn enables us to behave cleverly in new or changing situations and environments.

Although scaffolded by social contexts [89], the formation of abstract representations is often portrayed as an inherently individual cognitive process. Through development [90], training [5], experience [6] and exposure to language [91,92], an individual becomes increasingly aware of invariant features and relations in the environment motivating the formation of representations that group certain phenomena as tokens of the same type or an expression of the same relation. Christie & Gentner [90], for instance, demonstrate how children aged 3–4 gain increasing sensitivity to relational similarities. While younger children will typically point to the visual series ‘cat’-’cat’-’cat’ to be more similar to ‘squirrel’-’cat’-’cow’, than ‘elephant’-’elephant’-’elephant’, due to the object match (the fact that both series have a ‘cat’), older children are more inclined to recognize the structural similarity, in particular when the visual sets are presented side-by-side facilitating analogical mapping. Similarly, across a number of modalities and domains from visual perception and motor learning to reasoning and language, Raviv et al. [5] point to the way variability in individual learning experiences leads to better generalizations and facilitate transfer.

We argue that analogous, but socially distributed processes of abstraction unfold on the shorter time scale of situated collaborative problem solving. Rather than the product of generalization across the experiences of a single individual, here abstract representations emerge from the integration and generalizations across group members' varied contributions [93,94]. This happens when group members explicitly share, negotiate and dialogically reflect over their representations of a problem. Group members often bring slightly different intuitions, strategies and perspectives to the collaboration, and these discrepancies are instrumental in at least two respects: (i) by pooling their experiential contents, the group together represents more of the variation of the phenomenon than the individual group members; and (ii) the recognition of differences in individual representations can assist the group in recognizing higher-order, abstract similarities, for instance, a pattern or rule governing the varied stimuli.

In a series of experiments, Schwartz [13] thus compared the representations and solutions to a series of science-related problems generated by individuals working either alone or discussing in pairs. Across tasks, groups were found to derive more abstract representations of the problems, which in turn also facilitated their problem solving. For instance, groups were four times as likely to induce a parity rule to decide the direction of a set of interconnected gears than individuals working alone, who were more likely to represent the problem in terms of concrete physical causality (see also [95,96]).

Similarly, in an experiment by Voiklis & Corter [97], individuals and pairs were presented with a complex, rule-based categorization task. Based on combinations of their visual surface features, participants had to deduce the underlying categorization rules deciding if extraterrestrial aliens were dangerous or valuable. Pairs were found to have a performance advantage compared to individuals working alone. The authors suggest that these effects are contingent on the interactive nature of the pair condition, due to the way that ‘(…) negotiating reference compels collaborators to find communicable structure in their shared activity’ [97, p. 607]. The collaborative condition thus has two central elements: the presence of language, and the presence of interaction. Aspects of language itself have often been suggested to motivate abstract reasoning. Verbal labelling of categories has, for instance, been found to facilitate processes of categorization and abstraction [92,98]. The Voiklis and Corter study also included a ‘monologue condition’, where individuals were instructed to overtly use language in their categorization process. This, however, did not yield performance advantages pointing to the interactive, dialogical engagement as critical for the group advantage (see also [57]).

In a similar categorization task, Tylén et al. [14] compared the transfer of categories formed by groups and individuals during a training session to new and slightly differently looking stimuli in a test session. The study found that groups more successfully transferred and applied their training experiences to new items, in particular as categorization complexity increased. The authors suggest that, while individuals are inclined to form more detailed visual representations of stimuli from the training set, groups—through dialogical negotiation—derive more abstract category representations that are resilient to variation and thus transports more readily to new stimuli. Interestingly, variation in group performance was predicted by aspects of group members' dialogical coordination. Successful groups were characterized by greater diversity in the lexical contributions to the dialogue particularly during early training. This could suggest that groups that incorporate and integrate more variability in experience, perspective or strategies form more abstract representations (see also [99]).

The relation between properties of interaction and abstraction is also addressed by Mills & Redeker [100]. Participants performed a collaborative task assisting each other through a virtual maze. The task required them to converge on a vocabulary for talking about their whereabouts in the mazes. In previous studies using the maze game, participants would initially ground their references in local salient structures of the individual maze, while they only gradually come to realize and gravitate toward more efficient abstract ways of talking about the mazes (e.g. a coordination system of columns and rows) [101–104]. In Mills and Redeker's version, participants communicated through a chat interface which allowed the experimenters to seamlessly manipulate aspects of their dialogue [100]. In one condition, the interface would thus occasionally change participants' private turn-edits into public self-repairs which were transmitted to the partner. The authors found that groups who were presented with more self-repairs tended to jointly converge on more abstract representations of the maze environments. This points to the role of conversational grounding mechanisms in the convergence of referential expressions [105]. In particular, repair mechanisms seem to support interlocutors in detecting subtle misalignments of representation and can thus be instrumental in the establishment of abstract concepts for which there is less direct perceptual support [95,106].

(d) . Interim summary

Taken together, we have addressed the variability of task affordances and how they interact with collective cognitive processes. We suggest that social interaction stimulates different processes and serves different functions across task types. In contexts of convergent thinking, social interaction seems to facilitate access to and weight of information, while in contexts of divergent thinking, interaction stimulates exploratory search. We also find that social interaction stimulates processes of abstraction, by the negotiated integration and synthesis of variable individual information. In the following, we address how the properties of groups and qualities of interaction modulate collective cognitive processes. That is, we will briefly introduce the potential impact of group composition and the inherent variability within members in social groups, which can influence how various tasks or problems can be solved. In particular, we suggest that the outcome of some types of interactive tasks is affected by the level of group members’ cognitive diversity, such as tasks and problems involving processes of abstraction.

4. Collective processes shaped by group composition

As stated previously, social interaction can facilitate information integration and mutual adaptation to an extent that increases the potential of the group, and dialogical interaction seems to generally be a good strategy for reaching better solutions across many domains and tasks. On the other hand, this conveniently ignores the fact that social interaction is not a one-size-fits-all, and several observations have been made in terms of how properties of the task or the dynamics of the interaction itself modulates the collective achievements of groups. For instance, while dynamic interactions are often viewed as the paradigm case for increasingly better alignment and coordination between group members, individuals may also align too much.

As individuals interact and mutually adapt to each other, they are also subject to continuous social influence and thus become increasingly similar, that is, less unbiased and independent. As a result, the information being exchanged becomes correlated and redundant which can reduce the reliability of shared information [107–109]. In these cases, the benefit of social information integration and mutual adaptation may gradually decrease, limiting the potential of the group considerably. Indeed, in a collective perceptual decision-making task, Fusaroli & Tylén [22] found that the relative performance of groups was better predicted by the extent to which group members complemented each other in dialogue than the amount of alignment (see also [21]). Excessive alignment can thus lead to the phenomenon of groupthink [110,111], referring to a situation where the ‘thinking’ capacities of the group becomes no better than that of single individuals. Moreover, individuals may increasingly align their search to acquire similar information, which leads to less exploration of the space of possible solutions. For example, this is observed in relation to the passive emergence of filter bubbles and echo chambers [112,113], but also when individuals actively adapt too much to each other's knowledge and thinking [107,114].

In other words, if the group is composed of very similar individuals, or if group members in interaction align too much, we observe quantitative limits to the benefits of interaction, leading to information and innovation loss and suboptimal or irrational problem-solving strategies [115,116]. This is the case for both processes of convergent and divergent thinking alike. There is, to our knowledge, no studies yet that directly investigate groupthink and related phenomena in processes of abstraction, but there is evidence suggesting a relevant role of cognitive diversity [14]. In the following, we will address more generally how to potentially alleviate this challenge by considering the composition of groups.

(a) . The role of diversity in interactions and the process of abstraction

In many contexts of collaborations and group decisions, evidence suggests a positive role of introducing diversity. For example, cognitive search in contexts of problem solving can be made more explorative by increasing diversity, enabling individuals to search new sets of solutions [68]. Sometimes a distinction is made between functional and identity diversity [17,117]. Functional diversity relates to differences in the cognitive style, perspectives and strategies group members bring to the interaction, for example how to mentally approach a problem. However, groups might also benefit from identity diversity, which refers to differences in demographic factors and traits of individual members and their competences. Identity diversity can stimulate more flexible and novel thinking among group members by prompting individuals to revise their perspective and solutions in light of other diverse perspectives [118–120].

Introducing diversity in group interactions can potentially assist in de-correlating the information being exchanged and restoring reliability [56]. The assumption here is that individuals are sampling solutions from their different individual background knowledge, while also acquiring new information from sources that are not too similar. For example, introducing diversity in groups who had already increased the collective accuracy of wisdom-of-crowds from open discussion, was indicated to bring even larger benefits [74]. In the study, group members were re-sampled from other groups to cancel out previous experiences and convergence in a post hoc analysis. As a result, local correlations were broken, and average accuracy was increased. This is in line with previous literature suggesting that diversity of opinions is a central element in eliciting the wisdom-of-crowds effect [17,52].

Positive effects of diversity have also been observed in the domain of categorization, suggesting that processes of abstraction are affected by diversity. In an experiment, Fjaellingsdal et al. [99] presented participants with a complex rule-based categorization task. Analogous to the Wisconsin Card Sorting Test, every time participants reached high accuracy (10 correct trials in a row), the rule would change, and a new combination of stimulus features decided the category. However, unbeknown to the participants, a more abstract meta-rule governed the surface rule changes and participants had an advantage to the extent they uncovered this meta-rule. Participants were divided in three conditions: individuals, ‘similarity groups’ and ‘diversity groups’. Members of similarity groups were individually trained on the same categorization rule, while diversity group members were trained on different rules before entering a collaborative test phase.

The study found that diversity groups had a general performance advantage compared to individuals and similarity groups, and also seemed to recover faster from rule changes. The authors interpret these effects to suggest that the initial cognitive diversity induced by the training session, make diversity groups more likely to uncover the underlying abstract meta-rule, which in turn support more directed search when the categorization rules change. In other words, the dialogical sharing of diverse problem representations formed through individual training experiences make diversity group members susceptible to a more abstract level of similarity between the different training rules, which can be projected onto the changing rules of the test phase. Similar connections between cognitive diversity and processes of abstraction are suggested in Tylén et al. [14].

On the other hand, the role of diversity in interactions also has its limitations. We have previously addressed the importance of evaluating and weighing markers of the reliability of social information (see §3a). Additionally, interactions facilitate the necessary recalibration of such markers in order to enable better social information integration. However, interpreting signals of reliability in diverse others may be more challenging. For example, if individuals are too dissimilar (e.g. in competence) they may misinterpret reliability and erroneously weigh information accordingly [59,66]. Misinterpretation in such cases may in part be related to individual and cross-cultural differences in communication of confidence [121]. Similarly, diversity in cognitive style can in some tasks negatively affect consensus formation [20,122]. Also, in contexts of divergent thinking, high levels of diversity seem to be disruptive of joint search leading to a dispersed search style and low fluency [87].

The appropriate degree of diversity required in interactions is hard to estimate and will inevitably depend on characteristics of the group and the task. To summarize, group interaction research indicates that interactions are sensitive to the composition of the group: in order to communicate unambiguously and assess the social information appropriately, individuals benefit from the mutual adaptation and alignment of repeated dialogical interactions. At the same time, if the group is unable to facilitate some degree of diversity in terms of individual backgrounds and strategies of cognitive processing, it risks that the shared information evolves to be too correlated.

5. Concluding remarks

Humans are endowed with extraordinary capacities for dialogical sharing and negotiation of representations, which facilitate reasoning and problem solving across many domains and tasks. In this paper, we have investigated the intimate relation between social interaction and processes of information processing and problem solving. To situate these characteristics, we started from a conceptual distinction between convergent thinking, divergent thinking and abstraction characterizing different types of information processing. This tentative typology, as it were, assists us in observing how social interaction facilitates human reasoning and problem-solving processes in ways that are qualitatively different depending on the nature and affordances of the interaction.

In particular, we have argued that interaction can stimulate processes of abstraction when varied individual information is integrated and generalized across members of a group, facilitating the flexible transfer of experiences and knowledge between contexts. The current literature is sparse and investigations mainly address tasks affording convergent thinking (cf. [13,14,97,99]). However, we speculate that similar processes of collective abstraction might also facilitate divergent thinking. For instance, the formation of abstract representations may help escape cognitive fixedness, and in turn positively influence the flexibility and originality of task solutions [81].

Interestingly, abstraction processes might depend on the relative cognitive diversity of group members, for instance, in terms of alignment and information contribution [14,99]. However, more investigations on this relationship are warranted. To this end, we outlined how predictions of group advantages in collective thinking, or behaviour more broadly, are sensitive to the composition and relative diversity of members in a group. In particular, groups that converge to become too correlated may be subject to limited information, conformity and cognitive biases. Conversely, diversity can have consequences for the degree of mutual exploration of solutions and flexible thinking. For example, a recurring observation across tasks is that interaction seems to stimulate broader exploration of the solutions space. Similarly, the interactive integration of varied individual experiences and strategies appear to facilitate processes of abstraction and transfer learning. While many questions are still open, these initial observations suggest a critical role of interaction and diversity for processes of abstraction.

Data accessibility

This article has no additional data.

Authors' contributions

K.O.: conceptualization, writing—original draft, writing—review and editing; K.T.: conceptualization, writing—original draft, writing—review and editing.

All authors gave final approval for publication and agreed to be held accountable for the work performed therein.

Conflict of interest declaration

We declare we have no competing interests.

Funding

This work was supported by the European Union Horizon 2020, Twinning grant: TRAINCREASE (Agreement ID 952324).