M. Baker          (CNRS, France)

A. Blaye           (Universite de Provence a Aix, France)

C. O'Malley       (University of Nottingham, UK)

Abstract. For many years, theories of collaborative learning 

tended to focus on how individuals function in a group. More 

recently, the focus has shifted so that the group itself has become 

the unit of analysis. In terms of empirical research, the initial goal 

was to establish whether and under what circumstances 

collaborative learning was more effective than learning alone. 

Researchers controlled several independent variables (size of the 

group, composition of the group, nature of the task, communication 

media, and so on). However, these variables interacted with one 

another in a way that made it almost impossible to establish causal 

links between the conditions and the effects of collaboration. 

Hence, empirical studies have more recently started to focus less on 

establishing parameters for effective collaboration and more on 

trying to understand the role which such variables play in mediating 

interaction. In this chapter, we argue that this shift to a more 

process-oriented account requires new tools for analysing and 

modelling interactions.

1. Introduction

For many years, theories of collaborative learning tended to focus on how 

individuals function in a group. This reflected a position which was dominant 

both in cognitive psychology and in artificial intelligence in the 1970s and early 

1980s, where cognition was seen as a product of individual information 

processors, and where the context of social interaction was seen more as a 

background for individual activity than as a focus of research in itself. More 

recently, the group itself has become the unit of analysis and the focus has 

shifted to more emergent, socially constructed, properties of the interaction. 

In terms of empirical research, the initial goal was to establish whether and 

under what circumstances collaborative learning was more effective than 

learning alone. Researchers controlled several independent variables (size of the 

group, composition of the group, nature of the task, communication media, and 

so on). However, these variables interacted with one another in a way that made 

it almost impossible to establish causal links between the conditions and the 

effects of collaboration. Hence, empirical studies have more recently started to 

focus less on establishing parameters for effective collaboration and more on 

trying to understand the role which such variables play in mediating interaction. 

This shift to a more process-oriented account requires new tools for analysing 

and modelling interactions.

This chapter presents some of the major developments over recent years in this 

field, in both theoretical and empirical terms, and then considers the 

implications of such changes for tools and methods with which to observe and 

analyse interactions between learners. In so doing, we have tried to address 

both the work done in psychology and in distributed artificial intelligence 

(DAI). However, we have to acknowledge that this chapter has a bias towards 

psychology Ñ not only because it reflects the interests of the authors to a large 

extent, but also because DAI has focused more on cooperative problem solving 

than on collaborative learning.

At this point we need to make a brief comment on this distinction: learning 

versus problem solving and collaboration versus cooperation. While 

psychologists consider that learning and problem solving are similar processes, 

computer scientists still address them separately. Different research 

communities (DAI versus machine learning, for example) have developed 

different techniques, some for learning and some for problem solving. The 

'collaboration' versus 'cooperation' debate is more complex. Some people use 

these terms interchangeably. (Indeed, there is some disagreement amongst the 

authors themselves.) For the purposes of this chapter, in acknowledgement of 

distinctions that others in the field have made, we stick to a restricted definition 

of the terms. ÒCollaboration" is distinguished from "cooperation" in that 

cooperative work "... is accomplished by the division of labor among 

participants, as an activity where each person is responsible for a portion of the 

problem solving...", whereas collaboration involves the "... mutual engagement 

of participants in a coordinated effort to solve the problem together." (Roschelle 

&  Teasley, in press). 

Defining collaboration by the non-distribution of labour does not avoid 

ambiguities. Miyake has shown that some spontaneous division of labour may 

occur in collaboration: "The person who has more to say about the current topic 

takes the task-doer's role, while the other becomes an observer, monitoring the 

situation. The observer can contribute by criticising and giving topic-divergent 

motions, which are not the primary roles of the task-doer." (Miyake, 1986; p. 

174). O'Malley (1987) reported similar results with pairs attempting to 

understand the UNIX C-shell command interpreter. This distribution of roles 

depends on the nature of the task and may change frequently. For example, in 

computer-supported tasks, the participant who controls the mouse tends to be 

"executor", while the other is likely to be the "reflector" (Blaye, Light, Joiner, 

& Sheldon, 1991). Cooperation and collaboration do not differ in terms of 

whether or not the task is distributed, but by virtue of the way in which it is 

divided: in cooperation, the task is split (hierarchically) into independent 

subtasks; in collaboration, cognitive processes may be (heterarchically) divided 

into intertwined layers. In cooperation, coordination in only required when 

assembling partial results, while collaboration is "... a coordinated, 

synchronous activity that is the result of a continued attempt to construct and 

maintain a shared conception of a problem" (Roschelle &  Teasley, in press).

2. Theoretical Issues: the individual or the group as the unit

What is the nature of the dyad in collaborative learning? It can be viewed as 

comprising two relatively independent cognitive systems which exchange 

messages. It can also be viewed as a single cognitive system with is own 

properties. These two different answers to the question serve to anchor the two 

ends of the theoretical axis. At one end, the unit of analysis is the individual. 

The goal for research is to understand how one cognitive system is transformed 

by messages received from another. At the other end of the axis, the unit of 

analysis is the group. The challenge is to understand how these cognitive 

systems merge to produce a shared understanding of the problem. Along this 

axis, between the Ôindividual' and the 'group', we can find three different 

theoretical positions: socio-constructivist, socio-cultural and shared (or 

distributed) cognition approaches. 

In this chapter we talk about an ÔevolutionÕ along this axis because the social 

end has recently received more attention Ñ maybe because it has been 

previously neglected. We do not mean to imply than one viewpoint is better 

than another: scientists need both pictures from microscopes and pictures from 

satellites. Moreover, for the sake of exposition, the approaches will be 

presented as more different than they actually are. Both Piaget and Vygotsky 

acknowledge the intertwined social and individual aspects of development 

(Butterworth, 1982).

2.1. The socio-constructivist approach 

Although Piaget's theory focused mainly on individual aspects in cognitive 

development, it inspired a group of psychologists (the so-called ÒGenevan 

SchoolÓ) who in the 1970s undertook a systematic empirical investigation of 

how social interaction affects individual cognitive development (cf. Doise & 

Mugny, 1984). These researchers borrowed from the Piagetian perspective its 

structural framework and the major concepts which were used to account for 

development: conflict and the coordination of points of view (centrations). This 

new approach described itself as a socio-constructivist approach: it enhanced the 

role of inter-actions with others rather than actions themselves.

The main thesis of this approach is that "...it is above all through interacting 

with others, coordinating his/her approaches to reality with those of others, that 

the individual masters new approaches" (Doise, 1990, p.46). Individual 

cognitive development is seen as the result of a spiral of causality: a given level 

of individual development allows participation in certain social interactions 

which produce new individual states which, in turn, make possible more 

sophisticated social interaction, and so on.

Despite this theoretical claim, which suggests a complex intertwining between 

the social and the individual plane, the experimental paradigm used by its 

proponents involved two supposedly "individual" phases (pre- and post-test), 

separated by an intervention session in which subjects worked either alone 

(control condition) or in pairs. Evidence showed that, under certain conditions, 

peer interaction produced superior performances on individual post-test than 

individual training (for reviews, see Doise & Mugny, 1984; Blaye, 1988). The 

studies which established this tradition of research involved children in the age-

range 5-7 years, and relied essentially on Piagetian conservation tasks. Where 

working in pairs facilitated subsequent individual performance, the mediating 

process was characterised as "socio-cognitive conflict", i.e. conflict between 

different answers based on different centrations, embodied socially in the 

differing perspectives of the two subjects. The social dimension of the situation 

was seen as providing the impetus towards or catalyst for resolving the conflict. 

Such resolution could be achieved by transcending the different centrations to 

arrive at a more advanced "decentred" solution.

From this perspective, the question was asked: under which conditions might 

socio-cognitive conflict be induced? One answer was to pair children who were, 

from a Piagetian perspective, at different stages of cognitive development. 

However, it was emphasised that subsequent individual progress cannot be 

explained by one child simply modelling the other, more advanced, child. It has 

been repeatedly demonstrated that "two wrongs can make a right" (Glachan & 

Light, 1981). What is at stake here, then, is not imitation but a co-ordination of 

answers. Subjects at the same level of cognitive development but who enter the 

situation with different perspectives (due to spatial organisation, for instance) 

can also benefit from conflictual interactions (Mugny, Levy & Doise, 1978; 

Glachan & Light, 1982). 

Researchers in DAI report similar empirical results. Durfee et al (1989) showed 

that the performance of a network of problem solving agents is better when 

there is some inconsistency among the knowledge of each agent. Gasser (1991) 

pointed out the role of multiple representations and the need for mechanisms for 

reasoning among multiple representations (see Saitta, this volume). These 

findings concern the heterogeneity of a multi-agent system. Bird (1993) 

discriminates various forms of heterogeneity: when agents have different 

knowledge, use various knowledge representation schemes or use different 

reasoning mechanisms (induction, deduction, analogy, etc.). For Bird, 

heterogeneity is one of the three dimensions that define the design space for 

multi-agent systems. The other dimensions, distribution and autonomy, will be 

discussed later.

The success of the concept of conflict in computer systems is not surprising. 

This logical concept can be modelled in terms of knowledge or beliefs and 

integrated in truth maintenance systems or dialogue models. However, the main 

proponents of socio-cultural theory now admit that their view has probably been 

too mechanistic (Perret-Clermont et al., 1991). Blaye's empirical studies 

(Blaye, 1988) have highlighted the limits of "socio-cognitive conflict" as "the" 

underlying causal mechanism of social facilitation of cognitive development. 

Disagreement in itself seems to be less important than the fact that it generates 

communication between peer members (Blaye, 1988; Gilly, 1989). Bearison et 

al. (1986) reported that non-verbal disagreement (manifested for instance by 

moving the object positioned by the partner) was not predictive of post-test 

gains. 

The role of verbalisation may be to make explicit mutual regulation processes 

and thereby contribute to the internalisation of these regulation mechanisms by 

each partner (Blaye, 1988). This interpretation leads us to the socio-cultural 

theory discussed in the next section.

2.2. The socio-cultural approach

The second major theoretical influence comes from Vygotsky (1962, 1978) and 

researchers from the socio-cultural perspective (Wertsch, 1979, 1985, 1991; 

Rogoff, 1990). While the socio-cognitive approach focused on individual 

development in the context of social interaction, the socio-cultural approach 

focuses on the causal relationship between social interaction and individual 

cognitive change. The basic unit of analysis is social activity, from which 

individual mental functioning develops. Whereas a Piagetian approach sees 

social interaction as providing a catalyst for individual change, often dependent 

upon individual development, from a Vygotskian perspective, inter-

psychological processes are themselves internalised by the individuals involved. 

Vygotsky argued that development appears on two planes: first on the inter-

psychological, then on the intra-psychological. This is his Ògenetic law of 

cultural developmentÓ. Internalisation refers to the genetic link between the 

social and the inner planes. Social speech is used for interacting with others, 

inner speech is used to talk to ourselves, to reflect, to think. Inner speech serves 

the function of self-regulation.

A simple computational model of internalisation has been developed by 

Dillenbourg and Self (1992). The system includes two agents able to argue with 

each other. The agent's reasoning is implemented as an argumentation with 

itself (inner speech). Each learner stores the conversations conducted during 

collaborative problem solving and re-instantiates elements from the dialogue for 

its own reasoning. The learner may for instance discard an argument that has 

been previously refuted by its partner in a similar context. The psychological 

reality is of course more complex, what takes place at the inter-psychological 

level is not merely copied to the intra-psychological, but involves an active 

transformation by the individual. 

The mechanism through which participation in joint problem solving may 

change the understanding of a problem is referred to as ÒappropriationÓ 

(Rogoff, 1991). Appropriation is the socially-oriented version of Piaget's 

biologically-originated concept of assimilation (Newman, Griffin and Cole, 

1989). It is a mutual process: each partner gives meaning to the other's actions 

according to his or her own conceptual framework. Let us consider two 

persons, A and B, who solve a problem jointly. A performs the first action. B 

does the next one. B's action indicates to A how B interpreted A's first action. 

Fox (1987) reported that humans modify the meaning of their action 

retrospectively, according to the actions of others that follow it. From a 

computational viewpoint, this mechanism of appropriation requires a high level 

of opportunism from agent-B, which must integrate agent-A's contribution, 

even if this action was not part of his plans.

Like the previous approach, this theory also attaches significance to the degree 

of difference among co-learners. Vygotsky (1978) defined the Òzone of 

proximal developmentÓ as Ò...the distance between the actual developmental 

level as determined by independent problem solving and the level of potential 

development as determined through problem solving under adult guidance or in 

collaboration with more capable peers.Ó We will see that this concept is 

important to understand some empirical results.

Research in DAI does not directly refer to Vygotskian positions. This is 

somewhat surprising since the issue of regulation, which is central to the socio-

cultural theory, is also a major issue in DAI. In computational terms, regulation 

is more often referred to as a an issue of 'control' or 'autonomy'. For Bird 

(1993), it constitutes the second dimension of the design space for multi-agent 

systems. As in political structures, there exist centralised systems where control 

is achieved by a super-agent or a central data structure (e.g., blackboard 

architectures) and decentralised systems in which each agent has more 

autonomy. An agent is more autonomous if it executes local functions without 

interference with external operations (execution autonomy), if it chooses when 

and with whom it communicates (communication autonomy) and whether it 

self-organises into hierarchical, serial or parallel sub-processes (structural 

autonomy) (Bird, 1993). 

2.3. The shared cognition approach

The concept of shared cognition is deeply intertwined with the 'situated 

cognition' theory (Suchman, 1987; Lave, 1988 Ñ see also Mandl, this 

volume). For those researchers, the environment is an integral part of cognitive 

activity, and not merely a set of circumstances in which context-independent 

cognitive processes are performed. The environment includes a physical context 

and a social context. Under the influence of sociologists and anthropologists, 

the focus is placed largely on the social context, i.e. not only the temporary 

group of collaborators, but the social communities in which these collaborators 

participate. 

This approach offers a new perspective on the socio-cognitive and the socio-

cultural approaches, and has recently led to certain revisions by erstwhile 

proponents of the earlier theories. Perret-Clermont et al. (1991), for example, 

question the experimental settings they had previously used for developing the 

socio-constructivist approach. They noticed that their subjects tried to converge 

toward the experimenter's expectations. The subjects' answers were influenced 

by the meaning they had inferred from their social relationship with the 

experimenter. Wertsch (1991) makes similar criticisms against work in the 

socio-cultural tradition: social interactions are studied as if they occur outside a 

social structure. Through language, we acquire a culture which is specific to a 

community. For instance, we switch grammar and vocabulary rapidly between 

an academic seminar room and the changing rooms of a sports centre. But 

overall, beyond a vocabulary and a grammar, we acquire a structure of social 

meanings and relationships (Resnick, 1991) that are fundamental for future 

social interactions. 

This approach challenges the methodology used in many experiments where the 

subjects perform post-tests individually, often in a laboratory setting. More 

fundamentally, this approach questions the theoretical bases on which the 

previous ones rely: "... research paradigms built on supposedly clear 

distinctions between what is social and what is cognitive will have an inherent 

weakness, because the causality of social and cognitive processes is, at the very 

least, circular and is perhaps even more complex" (Perret-Clermont, Perret and 

Bell, 1991, p. 50). Collaboration is viewed as the process of building and 

maintaining a shared conception of a problem (Roschelle &  Teasley, in press). 

While the previous approaches were concerned with the inter-individual plane, 

the shared cognition approach focuses on the social plane, where emergent 

conceptions are analysed as a group product. For instance, it has been observed 

that providing explanations leads to improve knowledge (Webb 1991). From 

the 'individualist' perspective, this can be explained through the self-

explanation effect (Chi, Bassok, Lewis, Reimann & Glaser, 1989). From a 

'group' perspective, explanation is not something delivered by the explainer to 

the explainee. As we will see in section 5, it is instead constructed jointly by 

both partners trying to understand each other (Baker, 1991).

The idea that a group forms a single cognitive system may appear too 

metaphorical to a psychologist. It does not surprise a computer scientist. While 

the natural scale for a psychological agent is a human being, the scale of a 

computational agent is purely arbitrary. The (vague) concept of agent is used to 

represent sometimes a single neurone, a functional unit (e.g., the 'edge 

detector' agent), an individual or even the world. The granularity of a 

distributed system, i.e., the size of each agent, is a designer's choice. It is a 

variable that the designer can tune to grasp phenomena that are invisible at 

another scale. It supports systems with different layers of agents with various 

scales, wherein one may compare communication among agents at level N and 

communication among agents at level N+1. Dillenbourg and Self (1992) built a 

system in which the same procedures are used for dialogue among agents and 

for each agent's individual reasoning. Hutchins (1991) reports a two-layer 

system wherein he can tune communication patterns among the units of an agent 

(modelled as a network) and the communications among agents. According to 

the respective strengths of intra-network and inter-network links, he observes 

an increase or a decrease of the group confirmation bias which cannot be 

reduced to individualsÕ contributions. Gasser (1991) insists on properties of 

multi-agent systems which "will not be derivable or representable solely on the 

basis of properties of their component agents" (p. 112). 

3. Empirical Issues: effects, conditions and interactions

Not surprisingly, the different theoretical orientations we have just outlined 

have tended to employ rather different research paradigms. Generally, socio-

cognitive experiments concerned two subjects of approximately the same age 

(or the same developmental level) while the Vygotskian setting involved adult-

child pairs. Moreover, the Piagetian and Vygotskian paradigms used different 

collaborative tasks. We come back to these differences later. Other paradigms 

have been used independently of a particular theoretical framework, for instance 

the 'reciprocal teaching' paradigm (Palincsar and Brown, 1984; Palincsar, 

1987; Riggio et al., 1991) in which one learner plays the teacherÕs role for some 

of the time and then shift roles with the other learner. We can also distinguish 

empirical work according to the size of the groups involved (dyads versus 

larger groups) or ways in which mediating technologies are employed, as in 

computer-supported collaboration. 

There are also differences between the various approaches in terms of the 

research methods employed. In the socio-cognitive perspective, the 

methodology was to set up conditions hypothesised to facilitate learning and to 

compare the outcomes of this intervention with some control group. With such 

methods, collaboration is treated as a black box; the focus is on outcomes. In 

contrast, research from a socio-cultural point of view tends to employ micro 

genetic analyses of the social interaction. The focus is on the processes involved 

in social interaction. This is partly because of the importance attached to the 

concept of mediation in socio-cultural theory. Evidence is sought from dialogue 

for symbols and concepts which mediate social activity and which can in turn be 

subsequently found to mediate individual activity. The shared cognition 

approach obviously also favours the second methodology. 

Despite their intertwining, we have attempted to disentangle the different 

research paradigms and theoretical approaches. In what follows we describe the 

ÔevolutionÕ of empirical research within three paradigms that differ with respect 

to the number and the type of variables that are taken into account. 

3.1 The "effect" paradigm

Experiments conducted to answer the question Òis collaborative learning more 

efficient than learning alone?Ó were fairly straightforward. The independent 

variable was 'collaborative work' versus 'work alone'. The choice of the 

dependent measures varied according to what the investigators meant by 'more 

efficient'. The most frequent measure was the subjectÕs performance when 

solving alone the task they previously solved with somebody else. Some 

researchers decomposed this dependent variable into several other measures of 

performance, such as the improvement of monitoring and regulation skills 

(Brown & Palincsar, 1989; Blaye & Chambres, 1991) or a decrease in the 

confirmation bias. Within this paradigm, the precise analysis of effects is the 

only way to understand the mechanisms that make collaborative learning 

efficient.

This kind of research let to a body of contradictory results, within which the 

positive outcomes largely dominate (Slavin, 1983; Webb, 1991). Nevertheless, 

negative results cannot always be discarded as the result of experimental errors 

or noise. Some negative effects are stable and well documented, for instance the 

fact that low achievers progressively become passive when collaborating with 

high achievers (Salomon and Globerson,1989; Mulryan, 1992). There is a 

simple way to understand the controversial effects observed with the first 

paradigm: collaboration is in itself neither efficient or inefficient. Collaboration 

works under some conditions, and it is the aim of research to determine the 

conditions under which collaborative learning is efficient. This brings us to the 

second paradigm.

3.2 The "conditions" paradigm

To determine the conditions under which collaborative learning is efficient, one 

has to vary these conditions systematically. While the first experimental 

approach (in very general terms) varies only in terms of the dependent 

measures, the second experimental approach varies along two dimensions, both 

dependent and independent variables. Numerous independent variables have 

been studied. They concern the composition of the group, the features of the 

task, the context of collaboration and the medium available for communication. 

The composition of the group covers several other independent variables such 

as the number of members, their gender and the differences between 

participants. It is not possible here to give a complete overview of the findings 

concerning each of these variables. We will illustrate the work with three 

examples. 

3.2.1 Group heterogeneity

Group heterogeneity is probably the most studied variable. Scholars have 

considered differences with respect to general intellectual development, social 

status or domain expertise. They have considered objective and subjective 

differences in expertise (whether the subjects are actually different or just 

believe themselves to be so). We restrict ourselves here to objective differences 

between the task knowledge of each subject, a parameter which is relevant for 

DAI. For the socio-constructivist, this difference provides the conditions for 

generating socio-cognitive conflict. For the socio-cultural approach, it provides 

conditions for internalisation. However, the nature of differences differs within 

each theoretical approach. Socio-cognitive theory refers to symmetrical pairs 

(i.e., symmetrical with respect to general intellectual or developmental level) 

where members have different viewpoints, whilst socio-cultural theory is 

concerned with asymmetric pairs where members have different levels of skill. 

Piaget (1965) argued that interaction with adults leads to asymmetrical power 

relations or social status, and that in such interactions adults or more capable 

children are likely to dominate. The pressure to conform in the presence of 

someone with higher perceived status is not likely, in this view, to lead to 

genuine cognitive change. Nonetheless, Rogoff (1990) notes that many studies 

from a Piagetian perspective have involved pairing, for example, conservers 

with non-conservers. This is hardly pairing children of equal intellectual ability 

and is more consistent with the Vygotskian position. The point of difference 

between the two approaches then is not one of ÒequalÓ versus ÒunequalÓ pairs, 

but exactly what this equivalence entails. Researchers have attempted to 

determine the optimal degree of differences. If it is too small, it may fail to 

trigger interactions. If the difference is too large, there may be not interaction at 

all. For instance, in a classification task, Kuhn (1972) shown children solutions 

reflecting a difference of -1, 0 , +1 or +2 levels compared to their own 

solutions. He only observed significant improvement in the +1 condition. This 

notion of optimal difference also emerges in DAI where Gasser (1991) notes 

that agents need a common semantics even to decide that conflict exists! The 

'zone of proximal development' defines an optimal difference in an indirect 

way, i.e. not as a difference between subjects A and B, but as a difference 

between how A performs alone and how A performs with B's assistance.

Heterogeneity is also function of the size of the group. Empirical studies 

showed that pairs are more effective than larger groups, but heterogeneity is not 

the only factor that intervenes. Groups of three are less effective because they 

tend to be competitive, whilst pairs tend to be more cooperative (Trowbridge, 

1987). However, differences between group sizes seem to disappear when 

children are given the opportunity to interact with other in the class (Colbourn & 

Light, 1987).

3.2.2 Individual prerequisites

A second set of conditions defines some prerequisites to efficient collaboration. 

It seems that collaboration does not benefit an individual if he or she is below a 

certain developmental level. We consider here the absolute level of the 

individual, not his or her level relative to the other group members. According 

Piaget, for a conflictual interaction to give rise to progress, it must prompt 

individual cognitive restructuring. This implies that a resolution of conflict 

which would be exclusively based on social regulations (compliance from one 

partner for instance) would prevent interaction from being efficient. PiagetÕs 

theory predicts that pre-operational children lack the ability to decentre from 

their own perspective and therefore benefit from collaborative work. Indeed, as 

others have noted (Tudge & Rogoff, 1989), Piagetian theory in this respect 

leads to something of a paradox. It is not clear whether social interaction leads 

to the decentration necessary to benefit from collaboration, or that decentration 

has to happen before genuine collaboration can take place. Other research 

suggests that developmental factors need to be taken into account in resolving 

this issue. Azmitia (1988) looked at pairs of 5 year old with equivalent general 

abilities and found that when novices (with respect to the domain) were paired 

with experts on a model building task they improved significantly, whilst equal 

ability pairs did not. Azmitia argues that pre-schoolers may lack the skill to 

sustain discussions of alternative hypotheses. 

Vygotskian theory does not place the same sort of explicit developmental 

constraints on the ability to benefit from collaboration, but recent researchers 

(e.g., Wood et al., in press; Tomasello et al., 1993) have argued that certain 

skills in understanding other peopleÕs mental states are required for this which 

may set developmental constraints on collaborative learning. With a simple task 

this may be achievable at around 4 years of age, since children at this age can 

understand that another may lack the knowledge necessary to perform an action 

(or misrepresent the situation) and they can predict the state of the otherÕs 

knowledge. However, with more complex tasks, which demand reasoning 

using that knowledge to predict the partner's actions on the basis of their belief 

and intentions may not be achievable until about 6 years. In order to achieve 

shared understanding in a collaborative activity, the child must also be able to 

coordinate all these representations and have sufficient skills to communicate 

with respect to them. 

Research on peer tutoring has identified some conditions which are also relevant 

to collaborative learning. The first condition is that the child-tutor must be 

skilled at the task. Radziszewska and Rogoff (reported in Rogoff, 1990) found 

that training a 9 year old peer to the same level of performance as an adult on a 

planning task led to peer dyads performing as well as adult-child dyads and 

better than peer dyads in which neither partner had been trained. A second pre-

requisite is the ability of the child to reflect upon his or her own performance 

with respect to the task. Thirdly, in order to tutor contingently (i.e., to monitor 

the effects of previous help on subsequent actions by the learner), the child has 

to be able to assess whether the learnerÕs action was wrong with respect to the 

instructions or wrong with respect to the task, and then be able to produce the 

next tutorial action on the basis of both a representation of the previous 

instruction and an evaluation of the learnerÕs response to that instruction. Ellis 

and Rogoff (1982) found that 6 year old children were relatively unskilled at 

contingent instruction compared with adult tutors. Wood et al. (in press) found 

that 5 year old peer tutors were similarly unskilled relative to 7 year old tutors, 

and that 5 year olds tended to have difficulty inhibiting their own actions 

sufficiently to allow their ÒtuteeÓ to learn the task. However, children at this age 

were better ÒcollaboratorsÓ than 3 year old comparison dyads.

3.2.3 Task features

Tasks that have been typically used in collaborative learning from a Vygotskian 

perspective include skill acquisition, joint planning, categorisation and memory 

tasks. In contrast, the implication from socio-cognitive theory is that tasks 

should promote differences in perspectives or solutions. Typically, 

conservation and coordination tasks involve perspective-taking, planning and 

problem solving. There is thus little overlap in the nature of tasks investigated 

from the Piagetian and Vygotskian perspective. It is also clear that the nature of 

the task influences the results: one cannot observe conceptual change if the task 

is purely procedural and does not involve much understanding; reciprocally one 

cannot observe an improvement of regulation skills if the task requires no 

planning. Some tasks are less ÒshareableÓ than others. For instance, solving 

anagrams can hardly be done collaboratively because it involves perceptual 

processes which are not easy to verbalise (if they are open to introspection at 

all). In contrast, some tasks are inherently distributed, either geographically 

(e.g., two radar-agents, receiving different data about the same aeroplane), 

functionally (e.g., the pilot and the air traffic controller) or temporally (e.g., the 

take-off agent and the landing-agent) (Durfee et al., 1989).

3.2.5 Interactions between variables

Researchers rapidly discovered that the independent variables we have 

described so far do not have simple effects on learning outcomes but interact 

with each other in a complex way. Let us for instance examine the interaction 

between the composition of the pair and the task features. Studies that have 

compared the relative benefits of interacting with adults versus interacting with 

peers suggest that they vary according to the nature of the task, with peers being 

more useful than adults in tasks which require discussion of issues. Adult-child 

interaction may be more controlled by the adult rather than being a reciprocal 

relationship. Children are more likely to justify their assertions with peers than 

with adults. Rogoff (1990) notes that the differences between socio-cognitive 

and socio-cultural approaches with respect to composition of dyads are 

reconcilable. As she points out, whilst Vygotsky focused on acquiring 

understanding and skills, Piaget emphasised changes in perspectives or 

restructuring of concepts. Tutoring or guidance may be necessary for the 

former, whilst collaboration between peers of equivalent intellectual ability may 

be better in fostering the latter (Damon, 1984). So, how dyads or groups 

should be composed with respect to skills and abilities may depend upon what 

learning outcomes one is interested in (e.g., skill acquisition vs. conceptual 

change) and what tasks are involved (e.g., acquiring new knowledge versus 

restructuring existing knowledge). 

Although few studies have involved a direct comparison of peer collaboration 

and peer tutoring with the same task, the type of task may interact with the 

developmental level of the learner and the nature of the dyad. For example, 

Rogoff (1990) argues that planning tasks may be difficult for very young 

children because they require reference to things which are not in the Òhere-and-

nowÓ. However, adults may be able to carry out such metacognitive or 

metamnemonic roles that are beyond children, whilst demonstrating to the child 

how such processing could be accomplished. So, certain types of task may 

have inherent processing constraints which in turn place constraints on how the 

interaction should be supported.

3.3 The "interactions" paradigm

The complexity of the findings collected in the second paradigm led to the 

emergence of a third one. This introduces intermediate variables that describe 

the interactions that occur during collaboration. The question "under which 

conditions is collaborative learning efficient?" is split into two (hopefully 

simpler) sub-questions: which interactions occur under which conditions and 

what effects do these interactions have. The key is to find relevant intermediate 

variables, i.e., variables that describe the interactions and that can be empirically 

and theoretically related to the conditions of learning and to learning outcomes. 

This methodology however raises interpretation difficulties: if some types of 

interactions are positively correlated with task achievement, it may be that such 

interactions influence achievement or, conversely that high achievers are the 

only subjects able to engage these type of interaction (Webb, 1991). 

Nevertheless, underlying this approach is a fundamental shift: it may be time to 

stop looking for general effects of collaboration (e.g., in global developmental 

terms) and focus instead on more specific effects, paying attention to the more 

microgenetic features of the interaction. We will illustrate this viewpoint by two 

examples that are important both in psychology and in DAI: explanation and 

control.

3.3.1 Explanation

One way of describing interactions is to assess how elaborated is the help 

provided by one learner to the other. This level of elaboration can be considered 

as a continuum which goes from just giving the right answer to providing a 

detailed explanation. Webb (1991) performed a meta-analysis of the research 

conducted on this issue. This synthesis lead to two interesting results: 

elaborated explanations are not related to the explainee's performance, but they 

are positively correlated with the explainer's performance. Webb explains the 

first result by the fact that learning from receiving explanations is submitted to 

several conditions which may not be watched by the explainer, e.g., the fact the 

information must be delivered when the peer needs it, that the peer must 

understand it and must have the opportunity to us to solve the problem. The 

second result, the explainer's benefit, has been observed by other scholars 

(Bargh and Schul, 1980). Similar effects (called the self-explanation effect) 

have been observed when a learner is forced to explain an example to himself 

(Chi, Bassok, Lewis, Reimann & Glaser, 1989). A computational model of the 

self-explanation process have been proposed by VanLehn & Jones (1993). The 

main principle is that the instantiation of general knowledge with particular 

instances creates more specific knowledge, a mechanism that has also been 

studied in machine learning under the label 'explanation-based learning' 

(Mitchell et al., 1986). It would nevertheless be a mistake to consider self-

explanation and explanation to somebody else as identical mechanisms. This 

would dramatically underestimate the role that the receiver plays in the 

elaboration of the explanation. As we will see in section 5, an explanation is not 

a message simply delivered by one peer to the other, but the result of joint 

attempts to understand each other. Webb (1991) found that non-elaborated help 

(e.g., providing the answer) is not correlated with the explainer's performance 

and is negatively correlated with the explainee's performance in the case where 

the explainee actually asked for a more elaborated explanation. Webb explains 

these results by the fact that providing the answer while the student is expecting 

an explanation does not help him or her to understand the strategy, and may 

lead the explainee to infer an incorrect strategy or to lose his or her motivation to 

understand the strategy.

These findings partially answer the second sub-question of this paradigm, the 

relationship between categories of interaction and learning outcomes. The first 

sub-question concerns the conditions in which each category of interaction is 

more likely to occur. Webb (1991) reviewed several independent variables 

concerning group composition, namely, the gender of group members, their 

degree of introversion or extraversion and their absolute or relative expertise. 

With respect to the latter, explanations are more frequent when the group is 

moderately heterogeneous (high ability and medium ability students or medium 

ability and low ability students) and when the group is homogeneously 

composed of medium ability students. Some other group compositions are 

detrimental to the quality of explanations: homogeneous high ability students 

(because they assume they all know how to solve the problem), homogeneous 

low ability groups (because nobody can help) and heterogeneous groups 

comprising high, medium and low ability (because medium ability students 

seem to be almost excluded from interactions).

Verba and Winnykamen (1992) studied the relationship between categories of 

interactions and two independent variables: the general level of ability and the 

specific level of expertise. In pairs where the high ability child was the domain 

expert and the low ability child the novice, the interaction was characterised by 

tutoring or guidance from the high ability child. In pairs where the high ability 

child was the novice and the low ability child the expert, the interaction involved 

more collaboration and joint construction.

3.3.2 Control

Rogoff (1990, 1991) conducted various experiments in which children solved a 

spatial planning task with adults or with more skilled peers. She measured the 

performance of children in a post-test performed without help. Overall she 

found better results with adult-child than with child-child pairs but, more 

interestingly, she identified an intermediate variable which explains these 

variations. Effective adults involved the child in an explicit decision making 

process, while skilled peers tended to dominate the decision making. This was 

confirmed by the children who collaborated with an adult; those who scored 

better in the post-test were those for which the adults made the problem solving 

strategy explicit. These results are slightly biased by the fact that the proposed 

task (planning) is typically a task in which metaknowledge plays the central 

role. A socio-cultural interpretation would be that the explication of the problem 

solving strategy provides the opportunity to observe and potentially internalise 

the partner's strategy. From a socially shared cognition viewpoint, one could 

say that making the strategy explicit is the only way to participate in each other's 

strategy and progressively establish a joint strategy. 

4              Tools for observing interactions

When collaboration is mediated via a computer system, the design of this 

system impacts on the collaborative process. This mediation has methodological 

advantages: the experimenter may have explicit control over some aspects of 

collaboration (e.g., setting rules for turn taking, determining the division of 

labour or distribution of activities). The effects of the computer as medium also 

has pedagogical aspects: to support the type of interactions that are expected to 

promote learning. We describe three settings in which the computer influences 

collaboration..

4.1 Two human users collaborate on a computer-based task

Until relatively recently, one of the main advantages associated with computer 

use in schools was seen in terms of the potential for individualised learning. 

However, since schools generally have more students than computers, children 

often work in groups at the computer. Several empirical results suggest that 

group work Ñ at least dyadic work Ñ at the computer may enhance the benefit 

derived from the collaborative learning situation (for a review, see Blaye et al, 

1990). The specific questions to be addressed here deal with the extent to which 

learner(s)-computer interaction and human-human interaction can reciprocally 

enhance one another. For instance, interfaces which induce a specific 

distribution of roles between learning partners help to foster social interaction 

(O'Malley, 1992; Blaye et al., 1991). Such interfaces can serve to scaffold the 

executive and regulative aspects of the collaborative task. Another interesting 

example concerns the principle of immediate feedback which was seen as a 

critical feature in the first generation of educational software. It seems that 

immediate feedback may prevent fruitful exchanges between human co-learners 

because they then rely on the system to test their hypotheses instead of 

developing arguments to convince one another (Fraisse, 1987). In other words, 

aspects of the software can modify the socio-cognitive dynamics between the 

learning partners. In particular, the computerised learning environment 

constitutes in itself a mediational resource which can contribute to create a 

shared referent between the social partners (Roschelle & Teasley, in press). 

This research does not aim to build a Ôtheory' of human-human collaboration at 

the computer. The fact that the medium (i.e., the computer) is similar is by no 

means a sufficient reason to unify this field of research. Different interfaces, 

different computer-based tasks and activities may yield very different 

interactions and learning outcomes. However, for the sake of simplicity, we 

refer generically to computer-based activities in order to discuss the other 

general parameters which exert an influence (e.g., frequency of feedback, 

representations induced by the interface, role distribution, etc.).

4.2 Computer-mediated collaboration

While the previous setting was influenced by research in educational 

technology, the setting considered here has developed in parallel with work on 

'computer-supported cooperative work' (CSCW). This discipline covers 

communication systems from simple electronic mail to more advanced 

ÔgroupwareÕ (Shrage, 1992). There are various ways in which computers can 

support communication. In the past, this technology has been restricted to 

textual communication, but developments in broad bandwidth technology allow 

for more exciting possibilities such as synchronous shared workspaces and 

two-way audio-visual communication. Generally speaking, broad bandwidth is 

expected to afford greater opportunities for collaboration. This does mean that 

older technologies should be superseded. For instance, asynchronous text-

based communication provides time for reflection on messages and allows 

students lacking in confidence to learn nevertheless by ÒeavesdroppingÓ on 

conversations. In addition, low bandwidth communication may have some 

advantage in that, if it takes time and costs money in terms of connect time and 

if displays are restricted to a screen at a time, students may be forced to consider 

their responses more carefully.

Computer-mediated communication settings enables the experimenter to 

consider the communication bandwidth as factor. For instance, Smith et al. 

(1991) observed that task distribution was easier with a larger bandwidth (i.e., 

when seeing each other via video instead of audio-only communication) and 

when the setting gave users the feeling of being side-by-side, through having a 

shred workspace. They also observed that establishing face-to-face contact 

seems to be important during reflection stages, e.g., when partners discuss their 

observations, hypotheses or strategies. This fits in with research on mediated 

communication which, in general, suggests that face-to-face communication is 

more effective than audio-only communication for tasks which involve elements 

of negotiation (see Short, Williams & Christie, 1976).

4.3 Human-computer collaborative learning

Human-computer collaboration refers to situations where the system and the 

human user share roughly the same set of actions. We don't include systems 

which support an asymmetric task distribution, as between a user and a word 

processor, for instance. We describe two types of system where some learning 

is supposed to result from collaborative activities: apprenticeship systems and 

learning environments. Most of these systems do not actually fully satisfy the 

symmetry criterion.

An apprenticeship system is an expert system that refines its knowledge base by 

watching a human expert solving problems. The human expert is actually more 

teaching the system than collaborating with him or her, but the techniques 

developed are relevant to collaborative learning. The expert's behaviour is 

recorded as an example and the system applies explanation-based learning 

(EBL) techniques to learn from this example. In ODISSEUS (Wilkins, 1988), 

the system attempts to explain each human action in order to improve the 

HERACLES-NEOMYCIN knowledge base. An explanation is a sequence of 

metarules that relate the observed action to the problem-solving goal. If 

ODISSEUS fails to produce the explanation, it tries to "repair" its knowledge 

base by relaxing the constraints on the explanation process. LEAP (Mitchell et 

al., 1990) applies a similar approach to the design of VLSI circuits. The user 

can reject the proposed solution and refine the circuit him or herself. In this 

case, LEAP attempts to create rules that relate a given problem description to the 

circuit specified by the expert-user. LEAP explains why the circuit works for 

the given input signal and then generalises the explanation to create the rule 

premises. The interesting aspect is that these systems attempt to acquire the 

metaknowledge used by an expert, a central issue in the Vygotskian approach. 

However these systems rely on EBL techniques which requires a complete 

theory of the domain. Human learners theories are rarely complete and 

consistent. Some research has been carried out to by-pass this problem by 

integrating EBL with analogical and inductive learning (Tecuci and Kodratoff, 

1990).

Not surprisingly, the idea of human-computer collaborative learning has also 

been applied to educational software. It has firstly been suggested as an 

alternative technique for student modelling (Self, 1986), then as an attempt to 

break the computer omniscience that dominates educational computing 

(Dillenbourg, 1992). An interesting issue concerns the necessity to have a 

plausible co-learner. Along the continuum of design choices, we can 

discriminate levels of 'sensitivity'. At the first level, we could imagine an 

ELIZA-like system which randomly asks questions in order to involve the 

learner in plausible collaborative activities. Second level systems include a co-

worker, i.e., an agent which solves problems during the interaction but which 

is not learning. For instance, the Integration Kid (Chan & Baskin, 1988) does 

not learn, but jumps (an the tutor's request) to the next pre-specified knowledge 

level. At the third level, we have a real co-learner, i.e., a learning algorithm 

whose outputs are determined by its activities with the world, including its 

interactions with the human learner (Dillenbourg & Self, 1992). This research 

has not yet produced enough empirical data to determine whether more sensitive 

systems are more efficient than less sensitive one.

Another interesting issue to be addressed here is that the phenomena observed 

in human-human collaboration are repeated in human-computer collaboration. 

Salomon (1990) raises an important point in terms of knowing whether human-

computer interaction has potential for internalisation similar to human-human 

conversations. He suggests (Salomon, 1988) that some graphic representations 

could have this potential. We observed (Dillenbourg, in press) that learners 

were not very 'tolerant' with the computer: firstly, they had difficulties in 

accepting that the computerised partner makes silly mistakes, then, when the 

computer was repeatedly wrong, they stopped making suggestions altogether. 

The advantage of human-computer collaborative systems for the study of 

collaboration is that the experimenter can tune several parameters regarding to 

the pair composition (for instance, the initial knowledge of the co-partner). 

5 Tools for analysing interactions

At the present state of research, it is not clear which theoretical perspective is 

most fruitful for analysing interactions, although incidence of socio-cognitive 

conflict appears to be limited and restricted largely to Piagetian tasks (Blaye, 

1988). However, other researchers have shown that there are benefits in 

generating discussions of conflicting hypotheses for domains such as physics 

(e.g., Howe et al., 19??). A number of researchers (e.g., Webb, Ender & 

Lewis 1986; Blaye, Light, Joiner & Sheldon 1991; Behrend & Resnick 1989) 

have shown that various interactive measures other than "conflict" have a 

positive correlation with learning outcomes. It may be, as Mandl & Renkl 

(1992) suggest, that this uncertainty in the field is due to the fact that the 

Piagetian and Vygotskian perspectives as they stand are simply too global to 

allow proper explanation of the different results. These authors thus argue that 

"more local", domain/task-specific theories should be developed. As Barbieri & 

Light (1992) point out, "[s]tudies in collaborative learning at the computer 

usually do not go into a detailed analysis of interaction É" (p. 200), despite the 

fact that it is "É important to analyse the quality of the interaction more 

closely." (p. 200).

5.1 Analysis categories

Most researchers have generally used quite global categories of analysis 

grouped according to (at least) the following 'oppositions' : (1) social / 

cognitive, (2) cognitive / metacognitive, and (3) task / communicat