PAGER: Why AI Still Can't Be a Top CAD Engineer — and the Fix for the Semantic-Execution Gap

Forum topic · QianXun · 2026-05-19

Summary

A Shanghai AI Lab-led team introduces PAGER (Bridging the Semantic-Execution Gap in Point-Precise Geometric GUI Control), a framework that tackles why AI agents excel at reasoning about geometric tasks but fail at pixel-precise execution. On GeoGebra tasks, top models know the correct action ~88% of the time yet complete under 6% of tasks, because single-pixel errors cascade through topologically dependent drawing steps. PAGER addresses this with two mechanisms: dependency-structured planning, which builds an explicit component dependency graph and enforces topological execution order, and precision-aligned reinforcement learning, which penalizes click distance from target pixels and verifies geometric consistency against mathematical ground truth. On the PAGE benchmark, task success rates reportedly jump from under 9% to 62% — a more than 4x improvement. The author also raises open concerns: inference cost for real-time industrial use, scalability to 3D CAD environments where dependencies grow exponentially, and robustness when human users perturb points mid-task. PAGER signals AI's shift from conversational reasoning toward precision agentic control for CAD, circuit design, and other high-precision GUI work.

Background

The paper discussed here is PAGER: Bridging the Semantic-Execution Gap in Point-Precise Geometric GUI Control (arXiv 2605.15963, May 2026), by Jingxuan Wei, Xi Bai, et al., from Shanghai AI Laboratory, UCAS, and partner institutions. Its focus: agentic control, computer vision, and reinforcement learning for GUI agents requiring pixel-level precision.

The Problem: One Pixel Away From Failure

For everyday GUI tasks, rough clicks suffice. But in professional geometry work, a single pixel of error can be fatal. When researchers asked frontier models (e.g., GPT-4o and specialized GUI agents) to draw figures in GeoGebra, a striking paradox emerged:

AI knows *what to do* with 88% accuracy (pick the right tool, click the right points).
Yet full task success is under 6%.

The cause is topological dependency: geometric constructions build on earlier elements. If the first circle's center is off by one pixel, the tangent line built on it breaks the logical chain — and errors cascade like dominoes. This is the Semantic-Execution Gap.

PAGER's Two-Pronged Approach

1. Dependency-Structured Planning

component dependency graph

2. Precision-Aligned Reinforcement Learning

Distance penalty: a click 2 pixels off halves the reward; 5 pixels off counts as failure.
Geometric consistency verification: after drawing, the system mathematically checks whether the result truly is, say, the tangent to those two points.

This combination of Euclidean distance and mathematical ground truth forces pixel-stable motor control over thousands of trials.

Results

On the PAGE benchmark, complex-task success rate rises from under 9% to 62% — more than a 4x improvement, suggesting AI may begin to handle precision-critical work like CAD drafting and circuit design.

Open Concerns

The author flags unresolved issues:

1. Compute and latency cost: maintaining the dependency graph with pixel-level calibration may be expensive; the paper discusses inference time only briefly, and real-time industrial use may strain servers. 2. The 3D topology problem: PAGER is demonstrated in 2D GeoGebra. In 3D modeling software, spatial dependencies grow exponentially — whether the approach collapses in higher dimensions is unknown. 3. Robustness to interference: if a human user manually moves a point mid-construction, it's unclear whether PAGER self-heals or enters a logic deadlock.

Takeaway

True intelligence requires not just ambition but mastery of fine detail. PAGER injects topological intuition into mechanical action, teaching agents that imprecise foundations make any logical edifice a sandcastle. It marks AI's evolution from "a poet who can chat" toward "a surgeon with a scalpel" — closing the one-pixel gap between theorizing and doing.