FROM INDIVIDUAL LEARNING TO COLLECTIVE COGNITIVE LEARNING

Abstract This article develops the notion of collective cognitive learning as a possible direction for the development of artificial intelligence systems. Our starting point is that human cognition is cumulative: it took shape in the context of linguistic communication, joint activity, and the retention of verified experience beyond individual memory. We distinguish three modes of operation of multi-agent systems: exchange of answers, collective solution of a single task, and collective learning, in which the results of interaction change the means available to the system for solving subsequent tasks. Experimental evidence on human group work and on systems based on large language models (LLMs) shows that interaction at the stage of generating solutions is accompanied by process losses: independent generation followed by pooling of results is, as a rule, no worse than discussion. We therefore propose to look for the benefit of a collective not in joint generation but in the verification, selection, and retention of results. We propose a cross-task model of process losses constructed by analogy with Steiner’s approach, an operational definition of collective cognitive learning, and a formal model of the associative structure as a graph of hypothesis links with generation operators, a life cycle, provenance, a degree of confidence, a verification type, and retained alternatives. We analyse separately which factors may allow a collective of agents to outperform a single agent with the same external memory, and we introduce a distinction between verification against an objective criterion and agreed assessment by independent assessors, which prevents agreement among agents from being equated with truth. The approach is compared with existing methods for accumulating agent experience and for retrieving information from external memory, and we formulate a condition under which prior collective structuring may reduce the number of attention operations. The work continues the authors’ previous article on the goal-execution-verification architecture in mixed human-AI teams, transferring it from the level of a single task to the level of a sequence of tasks. The work is theoretical: we distinguish established findings, propositions, assumptions, and hypotheses, formulate nine testable hypotheses, and offer the research community a detailed protocol for testing them experimentally, which the authors are unable to implement themselves.