Thanks to Yunzong Xu and Bhaskar Ray Chaudhuri for the invite!

It's also the tone of the PC. They keep changing the process mid-day, shortening the timeline for ACs, and instead of acknowledging this, blaming the ACs and sending repeatedly threatening emails for ha being one day late on a task that we had a half the original assigned time for

The tl;dr is that balanced/fair algorithms do not necessarily comes at a cost. When you take incentives around data and network effects into account, fairness (here, through representativeness) can come at no cost.

And the paper that started it all: arxiv.org/abs/2501.19294

The cointreau or the apricot? I found that every time I put cointreau in a cocktail, I should have used half.

How did it turn out

With this work, we aim to add nuance to the discourse that decentralization on its own may not be the solution---rather, centralized decision-making should be more fine-grained to go beyond naive metrics, and understanding diversity of backgrounds/signals that make up qualified candidates.

This creates unfairness where *equally qualified* candidates are treated disparately based on, for example, how recognizable their alma matter is.

The easy way out here is spectacularly bad: hire candidates that they fully understand/have more information about, rather than candidates that are riskier (i.e., a top student from a small, not well-known high school).

Our main insight is as follows: while decentralization is useful to score candidates beyond just generic characteristics/take into account how successful they are expected to be for a specific team, our mathematical model shows that such decentralized designer will always take the easy way out.

In stage 2, decentralized entities (professors in a university or specific teams in a company) make admissions/hiring decisions based not only on "quality" (again, which is imperfectly perceived), but also how their specific attributes contribute to the specific team they aim to join.

We propose a mathematical model of 2-stage admission/hiring that questions this conventional wisdom.

In stage 1, a centralized designer makes initial hiring/admissions decisions based solely on a candidate's perceived "quality" (a noisy signal of their competences).