Mechanism design asks: given a desired outcome, what rules of interaction produce it as an equilibrium? The classic approach is analytical — derive the mechanism from first principles using auction theory, contract theory, or implementation theory. Each problem class requires its own mathematical treatment. The solutions are exact but bespoke.
Thoma, Piliouras, and Marris (arXiv:2603.07705, 2026) make the equilibrium itself differentiable. Their differentiable equilibrium blocks (DEBs) compute Nash equilibria of a game as a forward pass and propagate gradients through the equilibrium as a backward pass. The mechanism designer specifies the desired outcome as a loss function. The game's rules are the parameters. Gradient descent adjusts the rules until the equilibrium of the resulting game matches the target.
A single neural network trained through this pipeline handles multiple problem instances: contract design, machine scheduling, inverse equilibrium problems, games with two to sixteen actions per player. The architecture is game-agnostic — the same differentiable block computes equilibria for any normal-form game, and the surrounding network learns the mapping from problem instance to game parameters.
The structural point: the analytical barrier in mechanism design was never the equilibrium concept — Nash equilibrium is well-defined. The barrier was inverting it: given an equilibrium, finding the game that produces it. This inversion is what differentiation through the equilibrium achieves. The forward pass (game to equilibrium) was always computable. The backward pass (equilibrium to game parameters) was the hard direction. Making the forward pass differentiable turns the hard direction into gradient descent — the same trick that made rendering differentiable for inverse graphics and physics simulation differentiable for control.