AutoDesign: Meta-Harness Optimization for Long-Horizon Agentic Design
AutoDesign is a framework that aligns with human design priors, where a meta-harness optimizer guides a code agent to recursively improve harness based on rollout feedback. It is evaluated on the academic paper-to-poster generation task using PosterBench, which comprises a 100-paper Main Track spanning five disciplines and PosterBench-mini, a shared 10-paper subset for controlled evaluation. On the PosterBench Main Track, AutoDesign achieves the highest score of 78.32, surpassing the closed-source commercial system Claude Design by 7.45 points. Across seven controlled code-agent-model configurations, integrating the learned DesignHarness consistently improves performance.
AutoDesign introduces a meta-harness optimization framework for long-horizon agentic design, focusing on transforming multimodal sources into structured media outputs. The framework uses a meta-harness optimizer to guide a code agent in recursively improving a harness based on rollout feedback. Evaluation on the PosterBench benchmark shows AutoDesign achieves a top score of 78.32, outperforming Claude Design by 7.45 points, and the learned DesignHarness consistently improves performance across seven code-agent-model configurations.
The framework treats design generation as a long-horizon agentic process centered on a model-harness system. A meta-harness optimizer iteratively refines the harness by leveraging rollout feedback, enabling recursive self-improvement. The evaluation uses PosterBench, a benchmark with 100 papers across five disciplines, and a controlled 10-paper subset. The reported score of 78.32 and consistent improvements across configurations suggest the learned harness generalizes across different code-agent-model setups.
AutoDesign demonstrates a shift from static design pipelines to self-improving agentic systems. By surpassing a closed-source commercial system (Claude Design) on a public benchmark, it signals that open research frameworks can compete with commercial offerings in specialized design tasks. The focus on paper-to-poster generation highlights a practical application in academic and content creation workflows.
The framework could reduce manual effort in design tasks such as poster generation, potentially lowering costs for academic institutions, conference organizers, and content creators. If the approach generalizes, it may enable automated generation of other structured media outputs, creating opportunities for new design automation tools or enhancing existing agentic platforms.
Next observable signals include whether the AutoDesign framework is applied to other multimodal transformation tasks beyond poster generation, whether the learned DesignHarness is released or integrated into existing agent frameworks, and whether commercial systems adopt similar meta-harness optimization approaches. Further validation on larger or more diverse benchmarks would strengthen claims of generality.