TreasureX Researchcs.AIAI Agents Learn Graphic Design Skills From User Traffic Without Retraining
Hongyang Du, Lan Yan, Christian Flores, Asim Kadav · September 18, 2026
Graphic design is hard to automate because success is subjective — there is no clean scoring function an AI can optimize against. A new arXiv paper proposes a workaround: keep the underlying model frozen and let it accumulate an external "procedural memory" of natural-language skills. The agent drives professional design software through 230+ tools, and its memory grows in two ways — widening by writing new procedures for recurring subtasks, and deepening by revising existing ones based on what worked and what failed. A replay gate blocks any edit that fixes one case but breaks a previously successful one.
Across five rounds using 1,406 real user briefs and 1,869 auto-graded trajectories, with no weight updates and no human labels, the skill bank expanded from 76 to 139 entries. Execution success on Claude-Sonnet-4 rose from 72.7% to 99.3%, and the agent beat its skill-less counterpart in 61.8% and 67.6% of head-to-head comparisons on two frontier models. Widening and deepening together outperformed either alone (58.5% win rate, p=0.025).
The economic implication: capability gains may come from better memory management rather than ever-larger training runs, potentially lowering the cost of deploying reliable design agents. The approach also generalizes to any domain with noisy, unverifiable feedback.
Source: arXiv paper 2609.22086 — https://arxiv.org/abs/2609.22086
Educational content only, generated with AI assistance from public research abstracts. Not investment advice. Thank you to arXiv for use of its open access interoperability.