Designer-RSI Pushes GenEval2 to 99.3% Via Procedural Memory

A frozen frontier model, given a growing notebook of natural-language skills, drove Claude-Sonnet-4's GenEval2 execution success from 72.7% to 99.3% across five rounds, with no weight updates and no human labels. That is the headline claim of Designer-RSI, a paper by Hongyang Du, Lan Yan, Christian Flores, and Asim Kadav.

The setup: an agent operates professional design software through more than 230 tools, while an external procedural memory "accumulates and refines reusable design procedures from experience." The bank starts with 76 documentation-derived skills and grows to 139. Widening adds procedures for recurring uncovered subtasks; deepening rewrites existing procedures against their own successful and failed executions. A "matched replay gate admits only changes that repair failures without regressing observed successes."

The evaluation ran five rounds over 1,406 real user briefs and 1,869 automatically graded trajectories. Against a no-skill agent across four specialized design benchmarks, the skilled version won 61.8% of matchups on Claude-Sonnet-4 and 67.6% on Claude-Opus-4.6.

On 200 held-out briefs from a user-traffic benchmark, widening alone reached a 49.4% win rate, deepening alone 48.6%, and the combination 58.5% (p = 0.025). The abstract's closing line: "Procedural memory offers a practical route to continual adaptation of agents under noisy, unverifiable feedback."

The abstract publishes no per-benchmark accuracy numbers and does not identify the four specialized design benchmarks by name. It lands amid a run of agent-tooling releases we have been tracking this week.