AutoDesign: Meta-Harness Optimization for Long-Horizon Agentic Design
AutoDesign reports a 78.32 PosterBench score, beating Claude Design by 7.45 points on paper-to-poster generation.
The paper frames design generation as a long-horizon agent loop, with a meta-harness optimizer improving the harness from rollout feedback. Its evaluation centers on PosterBench, a new benchmark built from 100 academic papers across five disciplines. The learned DesignHarness raised the average score across seven controlled code-agent setups from 54.99 to 67.39. The authors also report a 40-minute autonomous run with 253 tool calls, 11 editing turns, and under $3 in cost. ArXiv · AI/CL/LG's note
The paper frames design generation as a long-horizon agent loop, with a meta-harness optimizer improving the harness from rollout feedback. Its evaluation centers on PosterBench, a new benchmark built from 100 academic papers across five disciplines. The learned DesignHarness raised the average score across seven controlled code-agent setups from 54.99 to 67.39. The authors also report a 40-minute autonomous run with 253 tool calls, 11 editing turns, and under $3 in cost. ArXiv · AI/CL/LG's note
score 5