Meta-Harness Optimization / 2026
AutoDesign
A research system for improving the harness that turns source material into faithful, editable artifacts.
From one paper to a complete research communication suite: an editable poster, presentation slides, a research webpage, and a narrated conference video with synchronized subtitles.
Research access
Next / the optimization object
System state DesignHarness / observable / editable
The optimization object
Improve the system that makes the artifact.
Each rollout leaves more than a poster. It leaves source context, an executable artifact, render evidence, diagnostic traces, and a bounded candidate update to the reusable harness.
Accepted transition Evaluate the output. Repair one module. Retain the next system state.
Paper communication suite / research generalization
One paper. A complete communication suite.
AutoDesign's validated DesignHarness produces source-grounded academic posters through autonomous optimization and human-guided refinement. We further study how the same Meta-Harness Optimization framework generalizes to slides, research webpages, and narrated videos.
PaperPoster · Slides · Web · Narrated Video
Academic PosterEditable HTML poster
01 / Paper to poster
An academic poster, made for the conference floor.
This high-resolution academic poster organizes paper evidence into a clear visual hierarchy for live presentation.
- Use
- Conference presentation
- Artifact
- Editable HTML poster
- Status
- Representative editable academic poster
24 slidesFormal academic deck
02 / Paper to slides
A formal research talk, structured across twenty-four slides.
AutoDesign is organized as a research presentation from the problem and method to the evidence and implications.
- Use
- Conference talk
- Artifact
- Formal academic PDF
- Status
- Exploratory Meta-Harness generalization
ResponsiveExecutable HTML
03 / Paper to web
An editorial research page built for reading, not just viewing.
AutoDesign becomes a responsive long-form artifact with grounded figures, quantitative evidence, and verified project links.
- Use
- Research dissemination
- Artifact
- Responsive HTML and CSS
- Status
- Exploratory Meta-Harness generalization
6 minutesConference video
04 / Paper to video
A conference talk that moves from method to evidence.
The AutoDesign conference video combines generated narration, synchronized subtitles, paper figures, and evaluation results in a timed visual sequence.
- Use
- Narrated conference presentation
- Artifact
- 1920 × 1080 H.264 video
- Status
- Exploratory Meta-Harness generalization
Representative optimization trace
The artifact changes once. The harness keeps learning.
AutoDesign reads the evidence left by completed rollouts, finds failures that repeat, and turns them into bounded updates to the system behind future artifacts.
State 00Observed
Recurring signal
Weak hierarchy · uneven evidence · underfilled regionsRepresentative sequence reconstructed from the AutoDesign workflow and artifact traces. It explains the mechanism; reported benchmark results appear below.
Inside DesignHarness
A design system the model can operate.
DesignHarness is AutoDesign's optimized design system, evaluated for academic paper-to-poster generation. It grounds the Designer in the paper, executes an editable artifact, inspects the browser result, and routes localized repair.
Compact method record
Evidence follows the harness.
DesignHarness is AutoDesign’s validated instantiation for academic posters. The artifact remains inspectable while system-level evidence accumulates across rollouts.
PosterBench / reported evidence
Measured across source grounding, communication quality, and executable artifact reliability.
Three independent views
Harness benefit, final quality, and human preference.
01 / Harness transfer · 10-paper controlled subset
One DesignHarness, seven fixed configurations.
Every evaluated configuration improves when the optimized DesignHarness is attached.
02 / Final quality
PosterBench
- 01 AutoDesign78.32
- 02 Claude Design70.87
- 03 OpenDesign69.45
100-paper PosterBench main track · selected system results.
03 / Blind human evaluation
- 11
- reviewers
- 55.2–77.8
- 95% confidence interval
- 67.6%
- tie-adjusted vs. Claude Design
Future directions
AutoDesign is currently validated for academic-poster DesignHarnesses. We are extending the same Meta-Harness Optimization process to slides, research webpages, and narrated video, each with its own artifact contract and evaluation protocol.
Research resources