Leadership as Coordination Control: Behavioral Signatures and the Recovery-Advantage Boundary in Multi-Agent LLM Teams

A new study finds that adding a process-level coordinator to multi-agent large language model teams does not uniformly improve accuracy, aligning with team science predictions that leadership is contingent on specific conditions [1]. The paper, posted to arXiv on June 17, 2026, operationalizes three classical leadership styles—transactional, transformational, and situational—as explicit controllers over a shared action vocabulary that includes explore, revise, accept, and synthesize [1]. A matched controller using the same actions but governed by an arbitrary rule recovers no better than majority voting, indicating that the theory-derived rule, not the vocabulary itself, drives any observed effect [1]. Across four task regimes and three open-weight model families, no single controller dominated by accuracy [1]. Transactional control matched a shared round-0 vote on all 12 model-regime combinations to within 1.3 percentage points [1]. Gains appeared only on the one combination where the round-0 majority was unreliable: the llama-4-scout model on a social task, where situational control posted an 8-percentage-point improvement over a flat interaction baseline [1]. The researchers used behavioral signatures—majority lock-in, exploration, and recovery from an incorrect round-0 consensus—alongside per-action ablations to isolate controller effects [1]. They propose a recovery-advantage account, tested with four boundary probes, which holds that a controller outperforms plain interaction only where the round-0 majority is unreliable, the task is recoverable, and undirected interaction does not already repair the error [1]. These conditions map onto established contingency theory concepts, including leadership substitutes, path-goal redundancy, and the situational readiness gap [1]. The authors argue that the largely null accuracy result is what the theory predicts, not a failure of the controllers themselves [1]. The study frames process-level coordination control as a contingency to be measured and theory-mapped, rather than a leaderboard to be topped [1]. The work contributes to a broader effort to understand coordination in multi-agent systems. While the paper does not address applications outside language model teams, the contingency framework it tests parallels research in other domains where coordination mechanisms are evaluated under varying conditions [4].