← Lab

Can a small model learn UI design from visual feedback?

Result

After visual reinforcement learning, Qwen3.5-9B won 67.7% of blind model judgments against its untrained base. It won 47.9% against the base model’s best-of-four approach.

How often the trained model wins

  • Against supervised training alone88%

    95% interval: 80.2–94.8%

  • Against the untrained base67.7%

    95% interval: 57.3–77.1%

  • Against base best-of-four47.9%

    95% interval: 37.5–58.3%

50% is an even match. Whiskers show 95% intervals. Best-of-four generates and ranks four UIs per result. These are GPT-6.1 Sol judgments, not human preferences.

Before and after training

Alder plan comparison

Before training

Alder plan comparison, before training, mobile

After training

Alder plan comparison, after training, mobile

Rill checkout

Before training

Rill checkout, before training, mobile

After training

Rill checkout, after training, mobile

Birch settings

Before training

Birch settings, before training, mobile

After training

Birch settings, after training, mobile
Three of the largest visual judge-score gains, selected from different task families. Each pair uses the same brief and seed. These are highlights, not typical results; the trained outputs still miss some requirements.

Methodology

Start with Qwen3.5-9B and 205 example interfaces, then run 512 reinforcement-learning updates using screenshot feedback from GPT-6.1 Sol. Test on 48 unseen briefs with two seeds, rendering each UI at desktop and mobile sizes. Blindly compare the trained model against the untrained base, supervised training alone, and best-of-four sampling: 96 comparisons per control. Training and evaluation use the same judge family, so this tests adaptation to its visual preferences.

Fine-tuning dataset (205 interfaces)
← All findings