Story · arXiv
Designer-RSI: Evolving Procedural Memory from User Traffic for Agentic Graphic Design (arXiv)
paper · Story page
A frozen model works design software through 230-plus tools while a memory of natural-language skills grows from real user briefs, gated by replay checks that block any change regressing an observed success. Five rounds take GenEval2 execution success on Claude-Sonnet-4 from 72.7% to 99.3%.
In plain words
- Researchers improved a computer design assistant by teaching it reusable steps from real design requests.
- It saves and revises written instructions without changing the artificial intelligence system that follows them.
- Changes are kept only when they fix past failures without spoiling tasks that previously succeeded.
- In the reported test, successful task completion rose from 72.7% to 99.3%.
- For design software builders, the results show better task completion without changing the underlying artificial intelligence.
Appeared in
- Approve one operation, run another: the binding failure in shipped agent products
Sep 22, 2026 · in the sections
Subscribe
Get the brief in your inbox
Pick daily, weekly, or both. Nothing is gated either way: every issue is on the site and in the feeds.
- Weekdays at 8:45am IST, one lead story and 6 to 9 items.
- Sundays, an argued synthesis rather than a recap.
- One click to leave, and quiet days say so in the subject line.