Human-Computer Interaction Institute
School of Computer Science, Carnegie Mellon University


The Knowledge-Learning-Instruction (KLI) Framework:
Toward Bridging the Science-Practice Chasm to
Enhance Robust Student Learning

Kenneth R. Koedinger, Albert T. Corbett, Charles Perfetti
with the current and past members of the Pittsburgh Science of Learning Center

June 2010


Keywords: Computational modeling, cognitive modeling, instructional theory, machine learning, learning science, second language learning, mathematics learning, science learning, robust learning, learing theory, knowledge components

The volume of research on learning and instruction is enormous. Yet progress in improving educational outcomes has been slow at best. Many learning science results have not been translated into general practice and it appears that most that have been fielded have not yielded significant results in randomized control trials. Addressing the chasm between learning science and educational practice will require massive efforts from many constituencies, but one of these efforts is to develop a theoretical framework that permits a more systematic accumulation of the relevant research base.

A key piece in such a theoretical framework is the development of levels of analyses that are fine enough to be supported by cognitive science and cognitive neuroscience, but also at levels appropriate to guide the design of effective educational practices. An ideal scientific solution would be a small set of universal instructional principles that can be applied to produce efficient and robust student learning for any educational goal. This holy grail is likely unattainable both because effective instructional practices in one subject-matter domain, like science, are often not effective in another, like second language and because even within a domain, specific instructional goals and contexts add restrictions to the application of principles. Thus, our strategy is to strive for instructional principles with as much generality as possible, while recognizing that (and stipulating how) their effectiveness may be constrained by student and knowledge characteristics of the target course.

What are the features of knowledge that may so constrain? To answer that question, we have embarked on an effort to create taxonomy of kinds of knowledge with a focus on functional cognitive characteristics that may determine which learning processes are most likely to produce different kinds of knowledge. Cognitive science has identified a vast array of learning strategies, processes, and mechanisms and these are also in need of order. Similarly, the learning and educational sciences have produced an array of instructional principles and practice recommendations. Our theoretical framework builds on work in these fields. The Knowledge-Learning Instruction (KLI) Framework encompasses the three components that constitute its name. Although we focus in this paper on the knowledge taxonomy, we begin to outline taxonomies of learning processes and of instructional principles. We suggest "complexity" as a key organizing principle and the use of the time course of the application of a unit of knowledge, a learning process, or an instructional principle as operational method for empirically grounding this notion of complexity.

The purpose of the framework and the taxonomic efforts is not just collection and organization, but generation of new research questions and hypotheses. One hypothesis-generating frame involves the relationship between the complexity of components of knowledge targeted by a course and the complexity of the instructional principles that are likely to be most effective and efficient in enhancing robust student learning in that course. See Table 5. We hypothesize that an instructional principle at a particular level of complexity will be effective for knowledge components of similar complexity and those of greater complexity, but not for ones of less complexity. In other words, complex instructional strategies (intended to support complex learning processes involving deliberate reasoning and sense-making) are best used only for the most complex of knowledge component goals, but simpler instructional strategies (intended to support simpler learning processes like memory) are relevant and effective for knowledge component goals of all complexity. A specific instance of this hypothesis is that the moderately complex instructional principle of prompting students to self-explain (or asking deep questions) is effective and efficient for learning more complex knowledge components (e.g., principles, like electric field properties, in mathematics and science) but is not effective and efficient for learning less complex knowledge components (e.g., categorical decisions, like English articles, in second language learning). The complexity dimension does not collapse onto domain differences. Thus, second language learning contains complex as well as simpler knowledge components, and so do math and science.

Thus the KLI framework has properties we think are essential to making a bridge from research to education. Its analysis of learning in terms of multi-level knowledge components reveals complexities that allow generalizations across domains. These generalizations, in turn, support instructional principles of high generality that, when combined with instructional goals, allow practical suggestions about curricula and intervention decisions.

42 pages

Return to: SCS Technical Report Collection
School of Computer Science homepage

This page maintained by