All pages
Demos, explanations, experiments and the original records.
Every current page, including demos and detailed results. Older versions remain under History and archive.
The lab and reference pages
- About SkinDeep
- Decisions without a written reply
- Earlier combined results
- Earlier investment analysis
- Earlier investment response
- Finding where to click
- History
- Ideas to explore
- Learning what you like
- Related research
- Research FAQ
- Run a demo
- Run the code locally
- SkinDeep
- Stopping early
- Test results
- The original paper
- Use cases
- What comes next
Learning what you like
- 3D assets and objects
- A feed without repeated prompting
- Architecture and interiors
- Art and visual design
- Automotive design
- Dating through mutual preferences
- Does preference learning work?
- Engineering with constraints
- Fashion and outfit choices
- Game levels and difficulty
- Genetics research
- How the model learns
- Interface and layout choices
- Learn more than one favorite style
- Learning materials
- Liking more than one style
- Makeup and appearance previews
- Material research
- Molecule and drug research
- Music and playlists
- News and content curation
- Patient preferences in approved options
- Product recommendations
- Protein research
- Recipes and meal choices
- Stop an image that misses your taste
- Stories and interactive fiction
- Teach it what you like
- Video edits and effects
- Voice and narration style
- Which example should come next?
Finding where to click
- Assistive pointing
- Check changing interfaces
- Click, or hold the action?
- Find visible text in a screenshot
- Finding the right point
- Mark regions in an MRI
- Medical-image landmarks and measurements
- Point to an app control
- Read an image. Return an action.
- See where the model would click
- Suggest an image annotation
- Where would it click?
Decisions without a written reply
- A decision without a written reply
- A small model, one decision
- Check ECG recording quality
- Choose a game move
- Choose a receipt total
- Choose the receipt total
- ECG labels without a written report
- EEG sleep-stage labels
- Find names to redact
- Find text to redact
- From a decision to a finished task
- Give search a second check
- Give your search a second check
- Let a trained model choose the moves
- Live chat moderation
- Other decision tests
- Process a queue together
- Qwen browser execution check
- Rank search results
- Reuse instructions and group messages
- Reuse the instructions
- Reuse work inside Qwen
- Route customer requests
- Route or ask for help
- Search real research
- Sort an inbox
- Try 100 real messages
- Try a small task classifier
- Try Qwen in your browser
- Watch the maze test
Stopping early
- Cascade: test results
- Check Qwen before continuing
- Check whether more layers are needed
- Commands on a small device
- Compare ways to decide sooner
- Does stopping hold up on new messages?
- Get an LLM result sooner
- MRI reconstruction with fewer model steps
- One model, two chances to answer
- Recorded browser comparisons
- Route between small and large models
- Stop after two layers, or continue
- Stop an unsafe image before it finishes
- Teach useful answers to appear earlier
- The original 600-message test
- Try a tiny model first
- Ways to make a decision sooner
- Where should processing stop?
- Which parts tolerate smaller numbers?