⌁ Awesome AI4CAD Hub Papers · Datasets · Projects
中文 GitHub
← Back to papers

Wrong Design Intent Is Worse Than None: A Derangement-Control Diagnosis of Header Conditioning in CAD Program Completion

arXiv 2026 64.7 method, application

TLDR

Tests whether LLM reads design-intent header in CAD program generation; wrong headers harm performance, but derangement control shows effect requires learned mapping.

Reasoning

Strengths include rigorous experimental design with causal controls, multiple seeds, and pre-registered conditions. Weaknesses are limited to a single model and dataset, and the abstract cuts off before discussing limitations or broader applicability.

Read-first score

Read-first score 64.7, weighted from topical fit, citation, graph, method, reproducibility, and recency signals. Original total remains 86.

Recency 8%
100

Uses a gentle age decay so recent papers surface without erasing older foundations. 2026

Methodology quality 25%
90

Screens visible abstract and analysis fields for experiment, dataset, baseline, metric, and limitation evidence. markers=baseline,evaluation,experiment,metric

Topical relevance 42%
53.8

Uses existing LLM keyword relevance scores normalized to 0-100. AI for CAD,computer-aided design,neural CAD,generative CAD,parametric CAD,B-Rep,boundary representation,constructive solid geometry,CSG,sketch extrusion,CAD generation,CAD reconstruction,text-to-CAD,image-to-CAD,point cloud to CAD,CAD program

Reproducibility 25%
46

Screens links and visible text for paper, code, dataset, artifact, and repository signals. pdf=True; code=False; dataset=False; markers=checkpoint,code

Field roles

FrontierMethodology anchor

Rank sensitivity

Stability: volatile; rank range: 5.

Keyword Scores

parametric CAD
9
sketch extrusion
9
CAD generation
9
CAD program
9
computer-aided design
8
generative CAD
8
B-Rep
8
boundary representation
8
AI for CAD
7
neural CAD
6
text-to-CAD
5
constructive solid geometry
0
CSG
0
CAD reconstruction
0
image-to-CAD
0
point cloud to CAD
0

Deep Analysis

Innovations

  • Causal derangement control (shuffled headers) that isolates learned header-to-program mapping from marginal distribution shift
  • Independent evaluation metric via executable geometric assertions on the B-rep solid, breaking circularity with the regex-based header extractor
  • Pre-registered 3-seed experimental design crossing prefix length and header type to diagnose header conditioning

Methodology

LoRA fine-tuning of Qwen2.5-Coder-1.5B on CadQuery programs with a five-feature CADCON header, then evaluating with three seeds under a pre-registered {0%, 40%}-prefix × {correct, wrong, masked}-header matrix. Adherence is scored by geometric assertions on the generated B-rep solid, and a derangement model trained with shuffled ground-truth headers serves as causal control.

Key Results

A wrong header significantly reduces adherence below the no-header baseline for polygonal and thin intents, while the derangement control remains immune (interaction significant on 3/3 seeds, p ≤ 4.2×10^{-3}). The independent metric deflates the apparent benefit of a correct header from +0.21 to +0.02, revealing metric circularity.

Limitations

  • Circle and tall design intents are at a baseline generation floor (~0) for this checkpoint, making them uninformative for the contrast
  • The harm is regime-specific: at 0% prefix the unconditioned baseline cannot generate valid CAD, so findings apply only to conditional completion
  • Study limited to a single model (Qwen2.5-Coder-1.5B) and one CAD program domain

Resources

Paper

Tags