Curtis' Thesis Defense this Friday [1:00 PM, 01/16, GCS 502C]
2 views
Skip to first unread message
Devansh Gupta
Jan 12, 2026, 1:21:43 PM (7 days ago) Jan 12
to USC Theory Group, usc-t...@googlegroups.com, CS Theory Group
Dear all,
Wishing you all a very happy new year! I am happy to share that Curtis Bechtel will be defending his thesis on Jan 16th at 1:00 PM at GCS 502C. It’s a big milestone, and we’d love for you to join us for the defense. Please find the title and abstract attached below.
Best,
Devansh
Title: Incentivizing Efficient Delegation without Payments
Date and Time: Friday, Jan 16th from 1:00pm to 3:00pm
Location: Ginsburg Hall (GCS) 502C
Date and Time: Friday, Jan 16th from 1:00pm to 3:00pm
Location: Ginsburg Hall (GCS) 502C
Zoom Link: https://usc.zoom.us/j/99047777628?pwd=r7ya1rcFWyadablDlqor3gLCvdibcD.1
Abstract: In delegation problems, a principal wants to search through a stochastic space of feasible solutions for one maximizing their utility, but they lack the ability to conduct this search on their own. Instead, they must delegate this search problem to one or more untrusted agents with distinct utility functions. The principal is then faced with the problem of designing a mechanism that incentivizes agents to find and propose a solution maximizing their utility. Importantly, the principal's power is limited to announcing which feasible solutions they would accept or reject, so we don't allow the principal to offer direct transfers of value, either positive or negative, for any outcome. Despite this limitation, there often exist mechanisms under which the principal is guaranteed a constant-factor approximation of their first-best utility. In this work, we propose three broad approaches to modeling delegation problems that address different aspects of the problem: combinatorial search and solution constraints, additive costs for searching, and delegating to multiple agents. We then show how the principal can achieve competitive approximations for several variants of each of these approaches.
Abstract: In delegation problems, a principal wants to search through a stochastic space of feasible solutions for one maximizing their utility, but they lack the ability to conduct this search on their own. Instead, they must delegate this search problem to one or more untrusted agents with distinct utility functions. The principal is then faced with the problem of designing a mechanism that incentivizes agents to find and propose a solution maximizing their utility. Importantly, the principal's power is limited to announcing which feasible solutions they would accept or reject, so we don't allow the principal to offer direct transfers of value, either positive or negative, for any outcome. Despite this limitation, there often exist mechanisms under which the principal is guaranteed a constant-factor approximation of their first-best utility. In this work, we propose three broad approaches to modeling delegation problems that address different aspects of the problem: combinatorial search and solution constraints, additive costs for searching, and delegating to multiple agents. We then show how the principal can achieve competitive approximations for several variants of each of these approaches.
Reply all
Reply to author
Forward
0 new messages