BONNIE'S NOTES

Complementarity and recombination in interacting groups form a feedback loop with the fitness landscape for individual intelligence, after Wu et al. (2024)

Planted: | Last tended: | 3 min read

Overview

@2024wuGroup propose a feedback loop between the individual cognitive capacities of theory-of-mind and compositionality, and the group-level capacities of complementarity—the ability of an interacting group to flexibly adopt specialized roles while working toward a joint goal1—and recombination of existing solutions in group problem-solving.

Related notes:

Individuals groups

  • Theory-of-mind facilitates complementarity: Compared to other proposed social learning mechanisms, ToM facilitates greater group complementarity by enabling key features of human coordination like selective imitation via inference mechanisms and adaptive delegation of both familiar and novel roles. Specifically, ToM is used to track others’ distinct goals, skills, preferences, and beliefs, as well as anticipating actions.
  • Compositional representations facilitate more effective recombination: Compared to classical models of cultural evolution by random mutations (i.e., noisy variations) in copying, compositionality allows agents to more flexibly integrate public information with private representations, maintaining structural coherence as individual knowledge structures are swapped with socially inferred fragments.
  • Co-evolution of complementarity and recombination: Distribution of more diverse knowledge via complementarity increases the pool of abstract structures available for recombination; on the other hand, recombination increases the space of concepts and goals, hence affording greater complementarity of specializations. Thus, the distinct cognitive capacities of ToM and compositionality have interacting effects at the group level.

Groups individuals

  • Complementarity eases individual cognitive load: Agents need only maintain a subset of knowledge required for survival, allowing for deeper individual specialization.
    • open-question Further empirical work: does increased cognitive load in the form of ToM or degree of compositionality reduce individual specialization?
  • Shared role decompositions reduces transactive memory costs: Agents with the same representation of complementary roles (i.e., same decomposition of the task) can use ToM to track expertise of other agents and thus engage in “on-demand” or “asynchronous” processing to retrieve knowledge only when needed.
    • open-question Further empirical work: does instability in interaction partners lead to an impaired ability to exchange social information?
  • Agents trained in social environments generalize better to new tasks: Group recombination gives each agent a larger combinatorial space of solutions for individual problem-solving tasks.
    • open-question Further empirical work: does this effect scale with the degree of specialization and recombination in the training group?

Footnotes

  1. Complementarity is distinct from reciprocity of prosocial behaviors often studied in game theory—the latter requires actions to match exactly, while the former often requires actions to differ in coordinated ways. ↩

Graph View