Combinatorial Contracts Through Demand Types

A team of researchers from top institutions has published a breakthrough paper titled 'Combinatorial Contracts Through Demand Types' on arXiv. The work tackles a core problem in AI and economics: how a principal (like a company) can design an optimal linear contract to incentivize an agent (like an AI system) to choose the best subset of actions from a set of n possibilities. This 'combinatorial action model' is notoriously complex, as the optimal contract depends on counting 'critical values'—points where the agent's preferred action set changes.

The researchers' key innovation is establishing a geometric link to consumer theory's 'demand types.' This view transforms the problem: bounding critical values becomes counting regions pierced by a 'contract ray.' Using this, they define the new class of All Substitutes and Complements (ASC) functions. ASC strictly generalizes known classes like gross substitutes and supermodular functions, admits at most O(n²) critical values, and is conjectured to be the maximal class with this property.

Beyond the structural result, the team also developed a new technique for efficiently computing 'demand queries' using 'value queries' for succinct demand types. Combining these structural and algorithmic advances yields polynomial-time algorithms for designing optimal contracts under new, broader classes of reward functions. This solves cases where actions exhibit simultaneous substitute and complementary relationships, a scenario common in real-world multi-step AI tasks but previously computationally intractable.