Transforming multimodal sources into condensed, structured media outputs can be viewed as a long-horizon agentic process centered on a model-harness system. However, existing paradigms remain static and fail to align with human design priors or accumulate reusable experience through empirical exploration.
To address this, researchers from Hugging Face present AutoDesign, a framework where a meta-harness optimizer guides a code agent to recursively improve the harness based on rollout feedback. This enables the system to align with human design priors and self-improve over time.
To evaluate the framework, the team introduced PosterBench, a benchmark for academic paper-to-poster generation, comprising a 100-paper Main Track across five disciplines and a shared 10-paper subset (PosterBench-mini) for controlled evaluation. On the Main Track, AutoDesign achieved 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 improved performance, increasing the average PosterBench Score from 54.99 to 67.39 (+12.4%). In a fully autonomous long-horizon loop, AutoDesign executed 253 tool calls and 11 editing turns within 40 minutes for under $3, reaching average conference-poster quality in human evaluation.
A system-blind human study further demonstrated that AutoDesign achieved the highest human preference among evaluated systems. The team also provides a live demo at designanything.ai and encourages community contributions via issues or pull requests.