Liking more than one style
You can like two different things without liking everything in between.
Live linear versus nonlinear preference learning; this illustrates separated preferences without a generator mixture.
A simple model may blur different tastes into an average. That average might not be something you like at all.
We are testing richer models and separate preference groups. The same question applies to designs, faces, music and changing moods.
Why the simple model struggles
The linear-model limit
For p(z)=sigmoid(w·z+b), a high-score set is a half-space intersected with the latent box. It is convex. Searching that same function in a different coordinate order cannot create several disconnected preferred islands.
The previous explanation conflated distinct acceptable solutions with distinct modes. Different-looking outputs can share a score; that does not establish a multimodal preference function in latent coordinates.
Alternatives to compare
Nonlinear features, small neural heads, mixtures, and separate contextual profiles are candidates. P04 begins with two known preferred regions and a deliberately disliked middle, comparing a linear head with quadratic features. P07 studies changing preferences separately.
Report mode recovery and false positives between modes, not just overall accuracy on an imbalanced sample.
Synthetic findings · Technical study protocols · Archived note