Research results
Does preference learning work?
The calculations work. Whether the suggestions are useful still needs a human test.
- Rate examples
- Learn which settings matter
- Score or adjust another example
83 browser calculation checks passed. In 80 generated bounded-edit cases, the solution agreed with an independent numerical solver.
On made-up preferences, the model learns simple patterns. It can fail when someone likes two separated styles or when a high predicted score does not match their real taste.
We have not yet shown that people prefer its suggestions. The calculation tests verify the mechanics, not that human outcome.
Try your own ratings · See how learning works
Data and detailed tests
These are generated numerical fixtures and synthetic preference rules, not a human-rating dataset.
Actual sampler, contexts and constraints · Five-seed learning results · Numerical fixtures · Original tables and comparisons · Mathematics