ACE is an agentic canvas editor built on a \emph{hierarchical scene-graph}, aimed at automating multi‑slide presentation design. Legacy document formats expose only flat, absolutely positioned elements, forcing agents to recompute coordinates and often breaking layouts; moreover, design lacks a unique ground truth, so reference‑based diff metrics penalize valid but different outputs.
Key innovations of ACE include:
- Hierarchical scene‑graph: provides a tree‑structured organization of elements, avoiding repeated coordinate calculations.
- Specialized action space: incorporates 98 tools covering common layout, graphic, and animation operations.
- CARE content‑aware router: supplies the agent with only the relevant slice of each deck, achieving an average token reduction of about 89%.
- Self‑correction loop: driven by a \textbf{ground‑truth‑free instruction‑following (IF) judge} whose natural‑language critique is fed back as the next‑turn instruction, enabling closed‑loop improvement.
Results show that with a fixed backbone, a single‑turn scene‑graph editor already matches an internally iterative HTML pipeline. Adding self‑correction lifts ACE to an IF score of 4.23 on the full 94‑task benchmark, versus 3.81 for the baseline (significant difference $p{=}.010$), while running 1.75\times faster and reducing cost by roughly 44%. Visual quality (VQ) metrics are statistically indistinguishable, yet among 26 blind raters, 58.7% clearly prefer ACE, and 81% favor the self‑corrected output. Rankings remain consistent across three judge families, and out‑of‑loop judges retain about two‑thirds of the self‑correction gain, indicating limited circularity. In 66% of cases the process halts after one pass, and a strict‑peak rollback eliminates every observed regression.
Blogger's Review: ACE’s combination of a hierarchical scene‑graph and a self‑correction mechanism markedly improves the reliability and efficiency of automated presentation design, especially by cutting token usage and cost. It merits further exploration and integration into real‑world design platforms.