Abstract

The influence of network structure on collective problem-solving is a central focus in collective intelligence. However, the causal mechanisms linking structure to collective outcomes remain unexplored. To explore these, we utilize an agent-based model, the Potions Task, which operationalizes problem-solving as a combinatorial process suited to information-theoretic analysis. We examine information-based metrics at the level of agent pairs across networks, analyzing variation over time to determine how they predict problem-solving across network structures. We measure redundancy (or conversely, synergy) of solutions discovered by agents with respect to the net-work’s global knowledge. Building on findings that small-world networks support efficient problem-solving, our results uncover the underlying mechanism by which high-performing networks achieve efficiency, namely by balancing local redundancy with long-range synergy. Furthermore, we find that synergy consistently predicts group performance, including in network structures typically considered inefficient at complex tasks. Synergy in information processing, measured at both local and global levels, therefore mediates the effects of structure and can override them entirely, suggesting that information-processing synergy, rather than network structure itself, is the primary driver of group performance. By formalizing a causal framework for information processing in collectives, our study provides a mechanistic account of collective problem-solving beyond network structure alone.