New paper proves optimal leader strategies are either pure or unstable

A new paper on arXiv (2606.17415) by Kamil Bulinski, Lang White, and Hung Nguyen tackles the fragility of optimal strategies in leader-follower (Stackelberg) games. The authors formalize a notion of stability: a Strong Stackelberg Equilibrium (SSE) is unstable if the follower has an alternative best response that strictly reduces the leader's payoff. They then prove a generic dichotomy: when the leader's utility is sampled from a continuous distribution, with probability one the optimal commitment is either pure (stable) or mixed (unstable). When both players' utilities are generic, the result strengthens: the unique optimal commitment is either pure and stable or mixed and unstable.

This result complements the classic work of von Stengel and Zamir by showing that even when leader values coincide generically, the strategy-level prediction is fragile whenever optimality requires randomization. The authors apply this to Stackelberg satisfaction games, disproving a prior conjecture with counterexamples and identifying conditions under which it holds. For AI and multi-agent systems, this means that any mixed-strategy equilibrium in leader-follower dynamics is inherently vulnerable — a critical insight for designing robust strategic agents.