Test-Time Scaling for CAD Generation via Verifier-Free Consensus Selection
Verifier-free 3D CAD consensus selection improves geometric metrics by 8-10% using model agreement over geometry and topology.
Key Findings
Methodology
This paper introduces a verifier-free post-processing approach for 3D CAD generation. It samples N parametric CAD programs from an LLM, compiles them into 3D models, and evaluates their agreement based on geometric (point cloud Chamfer distance) and topological (Euler characteristic) metrics. The selection follows a minimum Bayes risk framework, choosing the model with the highest consensus. The method requires no training, operates directly on candidate pools, and significantly improves geometric accuracy, reducing Chamfer distance by 8-10% compared to verifier-based selection. Experiments on the CADPrompt benchmark demonstrate its effectiveness across multiple LLMs and prompt variants.
Key Results
- On the CADPrompt benchmark, geometric consensus reduces Chamfer distance by approximately 8%, improves IoU scores, and matches or exceeds the performance of verifier-based methods. Topological metrics show no significant difference from verifier methods, with Tcorr reaching 84.9%. The method performs robustly across candidate pool sizes, with diminishing returns beyond N=9. Across models like Gemma 12B and gpt-oss-20B, consensus consistently outperforms random selection, with improvements of 8-10% in geometric metrics.
Significance
This approach addresses the reliance on external verification models, offering a simple yet effective alternative based on model agreement. It enhances the reliability of AI-driven CAD generation, reducing errors and improving geometric and topological fidelity. The method's training-free nature facilitates easy integration into existing pipelines, accelerating automation in engineering design, manufacturing, and prototyping industries. It also opens new avenues for leveraging model consensus in other 3D generation tasks, promoting scalable, high-quality AI-assisted design workflows.
Technical Contribution
The core innovation lies in applying a consensus-based selection mechanism directly on properties of compiled 3D models, bypassing the need for learned evaluators. The method combines geometric point cloud distances with topological Euler characteristic comparisons, unified under a minimum Bayes risk decoding framework. This fusion of explicit model attributes provides a robust, interpretable, and computationally efficient selection criterion, representing a significant departure from prior reliance on external classifiers or heuristics.
Novelty
This is the first work to utilize explicit geometric and topological properties of 3D CAD models for consensus-based candidate selection without external verifiers. Unlike prior approaches that depend on vision-language models or learned reward functions, this method exploits inherent model attributes, enabling direct, explainable, and training-free selection. Its simplicity and effectiveness mark a novel contribution to AI-based CAD automation.
Limitations
- The method's effectiveness depends on the diversity and quality of the candidate pool; if all candidates share similar errors, consensus may reinforce these mistakes. Additionally, the coarse topology measure (Euler characteristic) may not capture detailed differences, limiting its discriminative power for complex models. In cases where models deviate significantly from the true shape, geometric distances may lose sensitivity, leading to suboptimal selection.
Future Work
Future research could incorporate local feature comparisons or multi-scale metrics to better distinguish fine details. Combining multiple properties, such as dimensions, manufacturability, or simulation outcomes, could further enhance selection robustness. Extending the framework to sequential refinement processes or integrating it with iterative optimization strategies may improve model quality in complex scenarios. Exploring learned or adaptive consensus criteria could also be promising.
AI Executive Summary
This study introduces a novel, verifier-free consensus selection method for 3D CAD generation, addressing limitations of existing verification-dependent approaches. Traditional CAD generation systems often rely on external models, such as vision-language judges, to select the best candidate from multiple samples, which can introduce ambiguity and computational overhead. The proposed approach leverages the inherent properties of the generated models themselves—specifically geometric point cloud similarity and topological Euler characteristics—to evaluate agreement among candidates.
The core idea is inspired by the principle of minimum Bayes risk decoding, where the candidate most consistent with others in the pool is chosen. By compiling N sampled parametric CAD programs into 3D models, the method computes pairwise geometric distances (using symmetric Chamfer distance over uniformly sampled point clouds) and topological differences (via Euler characteristic comparisons). The candidate with the lowest average dissimilarity is selected as the consensus model. This approach is training-free, model-agnostic, and computationally efficient, requiring only pairwise comparisons.
Extensive experiments on the CADPrompt benchmark demonstrate the method's effectiveness. Results show that geometric consensus reduces Chamfer distance by 8-10%, improves IoU, and matches the topological accuracy of verifier-based methods. Across various large language models, including Gemma 12B and gpt-oss-20B, the consensus approach consistently outperforms random selection, especially as the candidate pool size increases. The findings suggest that agreement-based selection can serve as a reliable, scalable alternative to complex verification systems.
The significance of this work lies in its simplicity and practicality. By eliminating the need for additional training or external evaluators, it streamlines the pipeline for AI-driven CAD design, making it more accessible and faster. Its applicability extends beyond CAD to other domains where model agreement can serve as a proxy for quality. Limitations include potential failure when all candidates share similar errors or when the topology measure is too coarse. Future directions involve integrating local feature comparisons, multi-property fusion, and sequential refinement strategies.
Overall, this research advances the field by demonstrating that model agreement, measured through explicit geometric and topological properties, can effectively guide the selection of high-quality 3D models, paving the way for more autonomous and reliable AI-assisted design workflows.
Deep Dive
Abstract
Large language models can write parametric CAD programs from a natural-language description (text-to-CAD generation), but a single sample is often wrong. Increasing test-time compute by sampling multiple candidates only helps if a good candidate can be identified, yet no ground-truth model is available at generation time. Existing systems often require a separate verifier, such as a vision-language judge, to select among candidates. We investigate whether the candidate pool itself provides enough signal for effective selection and a verifier-free alternative. We introduce 3D CAD consensus selection, hereafter consensus selection: sample $N$ parametric CAD programs, compile them to 3D models, and return the candidate that agrees most with the rest of the pool. The method is training-free and compatible with existing CAD agents. We investigate geometric and topological notions of agreement, each of which improves its corresponding evaluation metric. On the exact candidate pools of a state-of-the-art CAD generation method, geometric consensus improves all three geometric metrics over the method's verifier, while topological consensus matches it on topology. Across every tested LLM and prompt variant, geometric consensus also improves geometric accuracy over random selection from the same pool, reducing Chamfer distance by $1-10\%$.