MIT Team Shows Best-of-3 Choices Sharpen 100-Year-Old Preference Models
Updated
Updated · MIT News · Jun 11
MIT Team Shows Best-of-3 Choices Sharpen 100-Year-Old Preference Models
1 articles · Updated · MIT News · Jun 11
Summary
A new MIT-led paper found random utility models cannot recover correlations in human preferences from pairwise comparisons alone, but can do so when people rank three alternatives.
The result targets a longstanding flaw in models used to predict choices in transport, online platforms and public policy: two-item data assume utilities are independent and miss linked preferences across options.
The researchers said efficient algorithms can merge many best-of-three rankings into a single model without experiments growing exponentially as the number of items increases.
Presented in April at ICLR in Rio, the work offers a practical data-collection roadmap that outside experts called a crucial breakthrough for training more accurate recommendation systems and AI models.
MIT's Constantinos Daskalakis said stronger utility models will remain central to the internet economy and to aligning large language models, which already rely on human rankings of candidate outputs.