AutoDesign: Meta-Harness Optimization for Long-Horizon Agentic Design of Academic Posters

Forum topic · 小凯 · 2026-08-15

Summary

AutoDesign is a framework that frames multimodal-to-media transformation as a long-horizon agentic process centered on a model-harness system. A meta-harness optimizer guides a code agent to recursively improve the harness based on rollout feedback, aligning outputs with human design priors and accumulating reusable experience. To evaluate the approach, the authors introduce PosterBench, comprising a 100-paper Main Track spanning five disciplines plus a 10-paper PosterBench-mini subset for controlled evaluation, focused on the academic paper-to-poster generation task. On the Main Track, AutoDesign reaches a highest score of 78.32, exceeding the closed-source Claude Design system by 7.45 points. Across seven controlled code-agent-model configurations, integrating the learned DesignHarness raises the average PosterBench score from 54.99 to 67.39, a 12.4 percent gain. In a fully autonomous loop, the system executes 253 tool calls and 11 editing turns within 40 minutes for under $3, achieving average conference-poster quality in human evaluation and the highest human preference among all evaluated systems.

Overview

Field: Natural Language Processing (NLP), agentic design systems Authors: Yaxin Luo, Haobin Jiang, Jialv Zou, Xu Huang, Wenhao Yan, Haodong Li, Zhengrong Yue, Jing Li, Xiaofu Chen, Xiaohan Zhao, Jiacheng Liu, Jiacheng Cui, Zhiqiang Shen, Xiaotong Li Release Date: 2026-08-13 arXiv: 2608.13560

Problem and Motivation

Transforming multimodal sources into condensed and structured media outputs can be fundamentally conceptualized as a long-horizon agentic process centered on a model-harness system. An ideal harness should align with human design priors and accumulate reusable experience through empirical exploration to enable recursive self-improvement. Existing paradigms remain static and fall short of this capability.

Method: AutoDesign

AutoDesign is a framework in which a meta-harness optimizer guides a code agent to recursively improve the harness based on rollout feedback. The design loop:

Starts from a prior harness aligned with human design conventions.
Uses rollout feedback from generated artifacts to identify weak points.
Has the code agent modify the harness code to address those weaknesses.
Re-rolls and iterates, accumulating reusable design experience.

Benchmark: PosterBench

To instantiate and evaluate the framework, the authors focus on the academic paper-to-poster generation task and introduce PosterBench:

Main Track: 100 papers spanning five disciplines.
PosterBench-mini: a shared 10-paper subset for controlled evaluation.

Key Results

Main Track: AutoDesign achieves the highest score of 78.32, surpassing the closed-source commercial system Claude Design by 7.45 points.
Generalization: Across seven controlled code-agent-model configurations, integrating the learned DesignHarness consistently improves performance, raising the average PosterBench score from 54.99 to 67.39 (+12.4%).
Autonomous loop: In a fully autonomous long-horizon run, the system executes 253 tool calls and 11 editing turns within 40 minutes for under $3, reaching average conference-poster quality in human evaluation.
Human study: A system-blind human study shows AutoDesign achieves the highest human preference among all evaluated systems.

Significance

The work demonstrates that dynamic, feedback-driven harness optimization can outperform static, even closed-source commercial systems on a challenging long-horizon multimodal generation task, while keeping compute and cost low (sub-$3 per poster in autonomous mode).