CADReasoner: Iterative Program Editing for CAD Reverse Engineering
arXiv:2603.29847v1 Announce Type: cross Abstract: Computer-Aided Design (CAD) powers modern engineering, yet producing high-quality parts still demands substantial expert effort. Many AI systems tackle CAD reverse engineering, but most are single-pass and miss fine geometric details. In contrast, human engineers compare the input shape with the reconstruction and iteratively modify the design based on remaining discrepancies. Agent-based methods mimic this loop with frozen VLMs, but weak 3D grounding of current foundation models limits reliability and efficiency. We introduce CADReasoner, a model trained to iteratively refine its prediction using geometric discrepancy between the input and the predicted shape. The model outputs a runnable CadQuery Python program whose rendered mesh is fed
View PDF HTML (experimental)
Abstract:Computer-Aided Design (CAD) powers modern engineering, yet producing high-quality parts still demands substantial expert effort. Many AI systems tackle CAD reverse engineering, but most are single-pass and miss fine geometric details. In contrast, human engineers compare the input shape with the reconstruction and iteratively modify the design based on remaining discrepancies. Agent-based methods mimic this loop with frozen VLMs, but weak 3D grounding of current foundation models limits reliability and efficiency. We introduce CADReasoner, a model trained to iteratively refine its prediction using geometric discrepancy between the input and the predicted shape. The model outputs a runnable CadQuery Python program whose rendered mesh is fed back at the next step. CADReasoner fuses multi-view renders and point clouds as complementary modalities. To bridge the realism gap, we propose a scan-simulation protocol applied during both training and evaluation. Across DeepCAD, Fusion 360, and MCB benchmarks, CADReasoner attains state-of-the-art results on clean and scan-sim tracks.
Subjects:
Graphics (cs.GR); Computer Vision and Pattern Recognition (cs.CV); Human-Computer Interaction (cs.HC)
Cite as: arXiv:2603.29847 [cs.GR]
(or arXiv:2603.29847v1 [cs.GR] for this version)
https://doi.org/10.48550/arXiv.2603.29847
arXiv-issued DOI via DataCite (pending registration)
Submission history
From: Dmitrii Zhemchuzhnikov [view email] [v1] Wed, 18 Feb 2026 10:27:29 UTC (7,842 KB)
Sign in to highlight and annotate this article
Conversation starters
Daily AI Digest
Get the top 5 AI stories delivered to your inbox every morning.
More about
modelfoundation modelbenchmark
Show HN: AI tool to merge people from two photos into one realistic group photo
At a friend's gathering, someone mentioned wanting to add her late father into a family photo. I figured this would be trivial — modern image models are powerful, just send two photos and ask the AI to merge them. She said she'd tried, but the results were weird and unstable, and she didn't know how to describe what she wanted. I went home and tried it myself. With a well-written prompt and two good photos, it works. But real-world use cases aren't two good photos — it's a modern family photo plus a damaged old portrait, or two old photos from different decades. That's when things fall apart. I looked at existing tools. Most showcase merges between clean, well-lit, modern photos. Nobody was solving the hard version: mismatched eras, damaged sources, different poses, different formality lev
I Built an AI Agent That Watches the Market While I Sleep
I have a full-time job and no time to watch the stock market all day. But I still trade — mostly US tech stocks. Last year I made at least three bad decisions because I was too tired or too rushed to think clearly. So I built an AI agent to do the watching for me. The stack: OpenClaw as the agent framework, Exa for information gathering, and Milvus as a personal memory store. Total cost: about $20/month. The NVIDIA Moment On February 26th, NVIDIA reported Q4 earnings — revenue up 65% year-over-year. The stock dropped 5.5%. I didn't find out until the next morning. But when I checked my phone, there was already a message from my agent, sent the previous evening: NVDA earnings analysis: Revenue beat expectations, but the market is skeptical about AI capex sustainability. In similar past situ
Knowledge Map
Connected Articles — Knowledge Graph
This article is connected to other articles through shared AI topics and tags.
More in Models
Show HN: AI tool to merge people from two photos into one realistic group photo
At a friend's gathering, someone mentioned wanting to add her late father into a family photo. I figured this would be trivial — modern image models are powerful, just send two photos and ask the AI to merge them. She said she'd tried, but the results were weird and unstable, and she didn't know how to describe what she wanted. I went home and tried it myself. With a well-written prompt and two good photos, it works. But real-world use cases aren't two good photos — it's a modern family photo plus a damaged old portrait, or two old photos from different decades. That's when things fall apart. I looked at existing tools. Most showcase merges between clean, well-lit, modern photos. Nobody was solving the hard version: mismatched eras, damaged sources, different poses, different formality lev
Agentic PHPUnit output
I was made a aware of PAO . And while it think it is a good tool I think we can do better by making it more useful for an LLM. The package has options for PHPUnit, Pest and ParaTest. I'm only going to focus on PHPUnit, version 12 in particular. The setup PHPUnit has an option to add extensions . The best way to let PHPUnit know your extension is in the phpunit.xml file. class= "Tests\Extensions\AgentAwareOutputExtension" /> To detect when PHPunit is run inside an agent I used the shipfastlabs/agent-detector library (I saw it in PAO). This library uses well known config variables to detect multiple agents. Because I'm trying out Mistral Vibe now I added a new script to composer.json. "test:agent" : "AI_AGENT=1 vendor/bin/phpunit" While PAO uses json as output, I want to use markdown. From t
Discussion
Sign in to join the discussion
No comments yet — be the first to share your thoughts!