Research FAQ
What the experiments show, how the methods work, and what is still untested.
Learning preferences
Can it learn what I like?
The demo learns patterns in your ratings and uses them to score or adjust drawings. Its calculations and simple made-up preferences have been tested; we have not yet shown that people prefer its suggestions.
Does choosing examples carefully save ratings?
Not consistently in our tests. The demo’s mix of likely likes, uncertain examples and random examples performed worse with noisy ratings than random sampling; selecting uncertain examples did better on the simple test rules.
Can it learn different tastes or changing moods?
A single linear model struggles with separated tastes. A nonlinear feature model helped in a synthetic test; keeping separate models for explicitly named contexts also helped in a small alternating-context test. Neither demonstrates learning a person’s changing moods.
Does a higher model score mean a better suggestion?
No. A model can confidently prefer the wrong thing. We tested selection from the same candidate pool and retained cases where maximizing the score disagrees with the known preference; a person’s new rating is still the real check.
Decisions and speed
Does the model need to write an answer?
No. A small classifier can read its internal numbers and return a label directly. Returning a number instead of one trained token did not, by itself, show a reliable speed advantage.
How early can it stop, and does that save time?
The learned checker stopped at layers 6, 12, 18 or 24 depending on the message. In a 100-message exploratory run it skipped 33.75% of transformer blocks and used about 34% less time; fresh reliability validation is still missing.
Would a tiny model with a larger fallback be simpler?
It is another useful approach. The tiny model handled 51 of 100 messages without Qwen, roughly halving measured time; the cascade also introduced one toxic-message miss that Qwen avoided.
Does it follow changing rules or understand a live conversation?
Not reliably in the tests so far: opposite rules exposed failures even at full depth. We have not tested these decision classifiers on live conversations with audio or video context.
Clicks and actions
Can it return a click location directly?
Yes. A published pointer model found 25 of 30 targets using all 28 transformer layers. Coordinate early stopping is untested; missing-target rejection has only an eight-example check.
Can those decisions finish a task?
A separate small grid model completed 7 of 10 mazes after more training and a rule against wall moves. That is not a completed GUI task: the screenshot tests return points and do not execute clicks.
Scope and limits
Can preferences stay within hard requirements?
Explicit bounds can prevent a known forbidden edit; new conflict cases check that mechanism. A high preference score is not a safety guarantee, and personalized instruction-following has not been validated.
Are the ratings and representations private?
The browser demo keeps its ratings on your device. That does not prove an encoded image hides private information; we have not run a reconstruction or privacy-attack study.
Has this been tested on real generators, music or other applications?
Not in the current experiments. Generator transfer, music, feeds, mutual preferences and engineering designs are proposed applications; validating them needs task data, appropriate models and, where relevant, participating users.
Will the results hold on another model or device?
A separate implementation replay reproduced saved decisions and layer stopping, including a published runtime version. We have CPU timings and a full-depth browser demo, but no matched GPU/NPU comparison or evidence that the trained classifiers transfer to another model.