MIT: 'Power of Threes' — Three-item Rankings Reveal Correlations in Preference Models

Colleagues, quick AI update: MIT researchers show how to improve classic Random Utility Models (RUM) for preference prediction.

Briefly:
- The study proves pairwise comparisons cannot recover correlations between options.
- Correlations become identifiable when respondents rank three alternatives in order (or combine best‑of‑three with pairwise).
- The authors propose efficient algorithms with a practical experimental budget.

Why it matters: uncovering preference correlations improves recommendation accuracy and model training (including LLM alignment).

What steps would you suggest to deploy this data collection in products? #AI #machinelearning #recommendations #MIT