New Methods for Privacy-Preserving Project Selection in Community Budgeting

Participatory budgeting—a mechanism that empowers residents to propose and decide on public spending—faces a growing challenge: managing the sheer volume of project submissions. A new paper from arXiv addresses this bottleneck through algorithmic shortlisting that can filter candidates without compromising privacy.

The Problem#

As participatory budgeting initiatives expand, organizers struggle to review hundreds or thousands of proposals efficiently. A naive approach using large language models to rank projects based on historical success patterns runs into accuracy issues and risks exposing sensitive voting data.

A Vote-Based Solution#

Researchers designed an alternative pipeline that leverages anonymous voting history rather than raw project features. Their key finding: this vote-based approach allows modern large language models to achieve performance comparable to classical machine learning techniques—solving the accuracy gap while maintaining privacy safeguards.

What It Reveals#

The study demonstrates that voter preferences in participatory budgeting follow stable patterns. These consistencies are strong enough that algorithmic shortlisting can effectively approximate which projects will succeed, enabling organizers to create an initial selection tier without manual review of every submission.

Implications#

The work opens a pathway for democratic processes to scale algorithmically while respecting participant privacy—a tension that has long complicated the use of automation in civic contexts. The research suggests that with the right methodology, organizing platforms can handle volume without sacrificing either efficiency or data protection.