The question
If taste is personal, a shared judge cannot choose well for one person. How much better is a selector that learns from that person’s own ratings, and after how many?
The test
We used the Jester dataset, in which people rated up to 100 jokes. Of the 15,412 people with at least 90 ratings, 3,000 were test people and 12,412 formed the crowd. The crowd gave each joke an average rating and ten taste factors. For each test person we revealed some of their ratings, then made up to 30 offers of three jokes they had also rated and asked each selector to choose the one they liked most. Chance is 33%.
The shared selector uses the crowd’s average rating, which is what one shared judge does. The personal selector adds a regression of the person’s own ratings on the ten taste factors. A third selector uses only the person’s lean toward or away from the unusual.
What it buys
A shared selector picks the person’s favourite of three 46.6% of the time. After 20 of a person’s ratings the personal selector does 2.9 points better, after 40 5.1 points, and after 80 6.4 points. The intervals are about ±0.6 points, so every gain from 10 ratings up is clear of zero.
Pick-of-3 accuracy gained over a shared judge, in points
It matters more when the options are all good. With three jokes from the crowd’s top third, a shared selector drops to 39% and the personal gain at 80 ratings grows to 7.4 points. That fits the earlier finding that people agree on the extremes and not the middle. Weak ideas can be filtered by anyone, but choosing among good ones has to be personal.
A taste for the unusual is not the selector
The lean-only selector gains half a point or less, even after 80 ratings. Personal taste in these data lives in many directions at once, not mainly on one scale of unusualness. A person’s preferred distance from the usual may be one input to a selector. It is not the selector.
Limits
These are jokes, not architecture. The taste factors come from 12,000 other people rating the same 100 jokes, and a new design has no ratings from anyone else, so a real selector has to place it from its content and will probably do worse. There is also no retest data, so we cannot compare accuracy with how often people agree with themselves.
What others have built
Personalised reward models are an active field. PAL (Chen et al., 2025) combines what users share with what makes each one different, and localises a new user’s preferences from a few examples, on text and on text-to-image tasks. Variational preference learning (Poddar et al., 2024) infers a user-specific latent from that user’s comparisons. Both are aimed at language and image models, not at designs. They suggest the approach is sound. They do not show it on designs.
What it means
A personal selector is worth building, and its first gain is modest. Our target is to match a person at 80% of the rate at which they agree with themselves, within 40 picks. We have not built it for designs. The blind round will give the first real picks to learn from.
Sources
- Countless Lab, selector learning curve, code in
research/selector-curve, and personal lean inresearch/jester-personal-z. - Jester joke ratings, Ken Goldberg et al., UC Berkeley, 2001.
- Dai-Wei Chen et al., PAL: Sample-Efficient Personalized Reward Modeling for Pluralistic Alignment, 2025.
- Sriyash Poddar et al., Personalizing Reinforcement Learning from Human Feedback with Variational Preference Learning, 2024.