Learning-Augmented Mechanism Design with and without Money

Speaker(s)

Artem Tsikiridis

Affiliation

Technical University of Munich

Language of the talk

English

Date

April 16, 2026, 12:15 p.m.

Room

room 4060

Seminar

Seminar Algorithmic Economics

---

While the worst-case analysis of algorithms provides useful robust guarantees, it often leads to uninformative bounds or impossibility results that may not reflect real-world obstacles. The learning-augmented framework aims to overcome the limitations of worst-case analysis by leveraging predictions about the input, while maintaining robustness guarantees even when the predictions are inaccurate. A recent line of work has proposed studying settings involving strategic agents under this lens. The resulting area of "learning-augmented mechanism design" seeks to understand when and how predictions can help overcome the barriers imposed by incentive compatibility.

In this talk, we present our recent work across three settings. First, we consider assignment problems without money in the private graph model, where we design a family of strategyproof mechanisms that smoothly interpolate between consistency and robustness, and establish a tight separation between deterministic and randomized mechanisms for bipartite matching (EC 2024). Second, we study budget-feasible procurement, showing that while predictions help in the online setting with secretary arrivals, they cannot improve the approximation guarantee in the offline setting (SAGT 2025). Third, we discuss how reserve prices based on data lead towards efficient equilibria in first-price auctions with heterogeneous autobidding agents (SODA 2026). We conclude the talk with a discussion on using predictions as a prior-free informational alternative for mechanism design.

---