The Cost of Delegation

ICLR 2026 Conference Submission373 Authors

Keywords: theory, alignment, multi-agent reinforcement learning

TL;DR: You can’t captain every oar—so you hire. Alignment issues are delegation problems.

Abstract: We study reinforcement learning and alignment through the lens of hierarchical coordination, where a principal steers many delegates with partial views and coupled effects. Starting from nonlinear dynamics, we identify the Cost of Delegation as the performance gap between centralized and decentralized control, decomposed into delegation, coordination, information, and surrogate mismatch components. We bound CoD, show that information value is decision-theoretic, and discuss implications for modern systems. Our work provides a theoretical foundation and new perspective for designing robust, scalable multi-agent systems.

Primary Area: reinforcement learning

Submission Number: 373