New Voting Algorithm Enhances Ranking Accuracy

A recent paper in the Journal of Artificial Intelligence Research introduces a voting method that improves ranking reconstruction by incorporating voter predictions of how other participants will vote. The approach builds on the “surprisingly popular algorithm,” which has been used since 2018 by research teams at UC Berkeley, including Hal Varians’ group, to aggregate preferences over single options through Bayesian prediction elicitation. The authors—Hadi Hosseini, Debmalya Mandal, Nisarg Shah, and Kevin Shi—published their work on 2026/08/24 ¹.

The core idea is simple: each respondent not only states her own preference but also predicts the votes of a subset of peers. When many such predictions are collected, the system can infer a more accurate global ordering even when experts are in minority. While this strategy already outperforms classical aggregation techniques for selecting a correct option from a small set, extending it to full ranked elections presents significant combinatorial challenges. The number of possible orderings grows factorially with the number of alternatives, making direct application computationally infeasible.

To address these difficulties, the team proposes six distinct elicitation formats that vary the amount of additional information required from voters. By tailoring the format to the available data, the method balances practicality against accuracy. Experimental results demonstrate that even modest amounts of predictive information substantially boost ranking quality compared to baseline methods like Borda count or Condorcet voting.

In practice, the researchers ran crowdsourcing experiments on Amazon Mechanical Turk (MTurk) where participants were asked to vote on several sets of alternatives while providing prediction cues about other workers’ preferences. Participants using the new algorithm achieved higher correlation with ground truth rankings than those employing traditional approaches.

The paper’s contribution is twofold. First, it provides empirical evidence that prediction‑based voting can recover true rankings more accurately than classical aggregation rules when voter predictions are incorporated appropriately. Second, by offering a spectrum of elicitation designs, it offers practitioners flexibility in choosing an approach that fits their constraints on survey length and participant burden.

Experts note that this work represents a significant step toward practical rank recovery under incomplete information. They point out that previous research had largely focused on single‑option selection, whereas reconstructing full orderings remains a major open problem due to combinatorial explosion ¹. By demonstrating that moderate prediction data can mitigate these challenges, the authors pave the way for future systems that combine human judgment with machine inference to produce reliable collective decisions.