RL can train design taste into models — but the taste belongs to the judge, not you

simonguozirui · x · 2026-10-30

Jasper Lu published an experiment: can RL instill design taste into a model? Using 2,093 webdev tasks from Xiaomi's open-sourced MiMo-V2.6 RL environments (7,780 agentic environments total), he trained two versions of Qwen3.8-27B with different LLM judges.

Key findings:

Subjective domains lack verifiable rewards; the standard LLM-as-judge approach spots clearly bad designs but judges taste poorly.
RL can produce stable taste, but the model's taste ends up mirroring the judge's — not the author's, not the average user's, and definitely not a real designer's.
The post covers experiment setup, training dynamics, results, how to build a better judge, and whether taste generalizes.

Original post →

RL can train design taste into models — but the taste belongs to the judge, not you

More from coding & agent