CADENA reconstructs a 3D mesh as an editable parametric CAD program. Instead of emitting the whole program in one pass, it adds one operation at a time, executes the partial program, and compares the target against what has been built so far — so the model always sees what is still missing, the way a human engineer works.<br><a href=\"https://cdn-uploads.huggingface.co/production/uploads/67d5a331eab66ce9cb01bae4/nGYkbQGksTZ8pICLAMIDx.png\" rel=\"nofollow\"><img src=\"https://cdn-uploads.huggingface.co/production/uploads/67d5a331eab66ce9cb01bae4/nGYkbQGksTZ8pICLAMIDx.png\" alt=\"Screenshot 2026-08-04 at 08.02.16\"></a><br>We also release CADENA-Bench: 3396 real mechanical parts from three industrial corpora, de-duplicated and grouped into six part families, reported per family so a method's weaknesses stay visible instead of averaging away. It is hard in a way existing test sets are not — every learned baseline loses roughly half its score moving from DeepCAD to real parts.<br>CADENA outperforms prior methods on DeepCAD, Fusion 360, MCB and CADENA-Bench, and reaches 0.910 voxel IoU on BenchCAD's Vision2Code. Reinforcement learning against executed geometry improves accuracy and cuts the invalid rate to 0.9%, since an operation that fails to build earns no reward.<br>Weights (SFT and RL), the benchmark, and inference code are all released.</p>\n<p><a href=\"https://cdn-uploads.huggingface.co/production/uploads/67d5a331eab66ce9cb01bae4/MPznU0BOj9qljUtnrLC0m.png\" rel=\"nofollow\"><img src=\"https://cdn-uploads.huggingface.co/production/uploads/67d5a331eab66ce9cb01bae4/MPznU0BOj9qljUtnrLC0m.png\" alt=\"Screenshot 2026-08-04 at 08.02.50\"></a></p>\n","updatedAt":"2026-08-04T05:03:17.823Z","author":{"_id":"67d5a331eab66ce9cb01bae4","avatarUrl":"/avatars/3ed437c874889e8a6db66c3ef88a60c0.svg","fullname":"DMITRII ZHEMCHUZHNIKOV","name":"zhemchuzhnikov","type":"user","isPro":false,"isHf":false,"isHfAdmin":false,"isMod":false,"followerCount":2,"isUserFollowing":false}},"numEdits":0,"identifiedLanguage":{"language":"en","probability":0.8775609135627747},"editors":["zhemchuzhnikov"],"editorAvatarUrls":["/avatars/3ed437c874889e8a6db66c3ef88a60c0.svg"],"reactions":[],"isReport":false}}],"primaryEmailConfirmed":false,"paper":{"id":"2608.00799","authors":[{"_id":"6a717194ec5082b9f872cedf","name":"Soslan Kabisov","hidden":false},{"_id":"6a717194ec5082b9f872cee0","name":"Gennadiy Savrasov","hidden":false},{"_id":"6a717194ec5082b9f872cee1","name":"Maksim Elistratov","hidden":false},{"_id":"6a717194ec5082b9f872cee2","name":"Antonio Rodriguez","hidden":false},{"_id":"6a717194ec5082b9f872cee3","name":"Daniil Ignatiev","hidden":false},{"_id":"6a717194ec5082b9f872cee4","name":"Nikita Gavrilov","hidden":false},{"_id":"6a717194ec5082b9f872cee5","name":"Rustam Uzdenov","hidden":false},{"_id":"6a717194ec5082b9f872cee6","name":"Alexey I. Boyko","hidden":false},{"_id":"6a717194ec5082b9f872cee7","name":"Igor Pasechnik","hidden":false},{"_id":"6a717194ec5082b9f872cee8","name":"Anton Konushin","hidden":false},{"_id":"6a717194ec5082b9f872cee9","name":"Andrey Kuznetsov","hidden":false},{"_id":"6a717194ec5082b9f872ceea","name":"Dmitrii Zhemchuzhnikov","hidden":false}],"publishedAt":"2026-08-01T00:00:00.000Z","submittedOnDailyAt":"2026-08-04T00:00:00.000Z","title":"CADENA: Stepwise CAD Reverse Engineering","submittedOnDailyBy":{"_id":"67d5a331eab66ce9cb01bae4","avatarUrl":"/avatars/3ed437c874889e8a6db66c3ef88a60c0.svg","isPro":false,"fullname":"DMITRII ZHEMCHUZHNIKOV","user":"zhemchuzhnikov","type":"user","name":"zhemchuzhnikov"},"summary":"Computer-Aided Design (CAD) underpins modern engineering, yet converting existing shapes into editable models still demands substantial expert effort. Most AI systems emit the entire CAD program in a single pass, never inspecting the intermediate geometry. In contrast, human engineers build a part feature by feature, checking after each operation what remains to be modeled. We introduce CADENA (Spanish for \"chain\"), a model that reconstructs a 3D mesh as a parametric CAD program, growing its sequence of operations one at a time and comparing the target with the currently predicted geometry at every step. We also address the lack of benchmarks for evaluating reverse-engineering methods on mechanical parts, introducing CADENA-Bench, a benchmark that measures performance across categories of mechanical parts. CADENA outperforms prior methods on CADENA-Bench and on the DeepCAD, Fusion 360, and MCB datasets. Code is available at https://github.com/zhemdi/cadena, model weights at https://huggingface.co/kulibinai/cadena, and CADENA-Bench at https://huggingface.co/datasets/kulibinai/cadena-bench.","upvotes":13,"discussionId":"6a717194ec5082b9f872ceeb","githubRepo":"https://github.com/zhemdi/cadena","githubRepoAddedBy":"user","githubStars":1},"canReadDatabase":false,"canManagePapers":false,"canSubmit":false,"hasHfLevelAccess":false,"upvoted":false,"upvoters":[{"_id":"67d5a331eab66ce9cb01bae4","avatarUrl":"/avatars/3ed437c874889e8a6db66c3ef88a60c0.svg","isPro":false,"fullname":"DMITRII ZHEMCHUZHNIKOV","user":"zhemchuzhnikov","type":"user"},{"_id":"66fe582c7d722f08799fd0a3","avatarUrl":"/avatars/661f405b114729ac60dfc68444c6df0c.svg","isPro":false,"fullname":"Lana Green","user":"lana-green-lori11","type":"user"},{"_id":"6568b7402419be607254e197","avatarUrl":"/avatars/fc22e5ba9c50a4e12e7030a7ae490b1b.svg","isPro":false,"fullname":"Zadorozhny","user":"pavul","type":"user"},{"_id":"66680c6451545a8b46c6fd21","avatarUrl":"/avatars/03707f5ea4e2aa8dc825a9782b00ed85.svg","isPro":false,"fullname":"Aibek Alanov","user":"ai-alanov","type":"user"},{"_id":"643984dceb7c5616ef3f5d54","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/643984dceb7c5616ef3f5d54/10JRkblrRIEVci6UJwvPz.jpeg","isPro":false,"fullname":"Andrey Kuznetsov","user":"kuznetsoffandrey","type":"user"},{"_id":"6a718fdecbb33221e8c2614f","avatarUrl":"/avatars/06d926423da2efec9ca9c7bc99c43d12.svg","isPro":false,"fullname":"Gennadii","user":"GENBBB","type":"user"},{"_id":"661e44cf1d8ffc49b57ba07e","avatarUrl":"/avatars/3e937cc4f784b369b9f996ba82d1b81d.svg","isPro":false,"fullname":"Andrey Galichin","user":"andreuka18","type":"user"},{"_id":"6310ff34bc152fa3e810c186","avatarUrl":"/avatars/bfd63bcd81548283f5e496e3693bf143.svg","isPro":true,"fullname":"Elizaveta Goncharova","user":"Elizaveta","type":"user"},{"_id":"64eb597d34d4f8996629a7cb","avatarUrl":"/avatars/0bb50bfe22c7685f960b0bb5dfa22b45.svg","isPro":false,"fullname":"Aleksandr Lazarev","user":"SashaDance","type":"user"},{"_id":"6805de192ff5594f0d311c6f","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/6805de192ff5594f0d311c6f/1Fs4_a1fSTDEGiFaLKnDl.jpeg","isPro":false,"fullname":"Mike Shevchenko","user":"mikhashev","type":"user"},{"_id":"6760832b899f4731eb7eb6b7","avatarUrl":"/avatars/40297d3447cf045be1fba7c39683b3c8.svg","isPro":false,"fullname":"merray","user":"merray0","type":"user"},{"_id":"66d07487b86483cf13320564","avatarUrl":"/avatars/ad59d8edc680336e17a9a4f095a5d0da.svg","isPro":false,"fullname":"Ramil Khafizov","user":"smileyenot983","type":"user"}],"acceptLanguages":["en"],"dailyPaperRank":0,"markdownContentUrl":"https://huggingface.co/buckets/huggingchat/papers-content/resolve/2608/2608.00799.md","query":{}}">
CADENA: Stepwise CAD Reverse Engineering
Abstract
Computer-Aided Design (CAD) underpins modern engineering, yet converting existing shapes into editable models still demands substantial expert effort. Most AI systems emit the entire CAD program in a single pass, never inspecting the intermediate geometry. In contrast, human engineers build a part feature by feature, checking after each operation what remains to be modeled. We introduce CADENA (Spanish for "chain"), a model that reconstructs a 3D mesh as a parametric CAD program, growing its sequence of operations one at a time and comparing the target with the currently predicted geometry at every step. We also address the lack of benchmarks for evaluating reverse-engineering methods on mechanical parts, introducing CADENA-Bench, a benchmark that measures performance across categories of mechanical parts. CADENA outperforms prior methods on CADENA-Bench and on the DeepCAD, Fusion 360, and MCB datasets. Code is available at https://github.com/zhemdi/cadena, model weights at https://huggingface.co/kulibinai/cadena, and CADENA-Bench at https://huggingface.co/datasets/kulibinai/cadena-bench.
Community
CADENA reconstructs a 3D mesh as an editable parametric CAD program. Instead of emitting the whole program in one pass, it adds one operation at a time, executes the partial program, and compares the target against what has been built so far — so the model always sees what is still missing, the way a human engineer works.
We also release CADENA-Bench: 3396 real mechanical parts from three industrial corpora, de-duplicated and grouped into six part families, reported per family so a method's weaknesses stay visible instead of averaging away. It is hard in a way existing test sets are not — every learned baseline loses roughly half its score moving from DeepCAD to real parts.
CADENA outperforms prior methods on DeepCAD, Fusion 360, MCB and CADENA-Bench, and reaches 0.910 voxel IoU on BenchCAD's Vision2Code. Reinforcement learning against executed geometry improves accuracy and cuts the invalid rate to 0.9%, since an operation that fails to build earns no reward.
Weights (SFT and RL), the benchmark, and inference code are all released.
Upload images, audio, and videos by dragging in the text input, pasting, or clicking here.
Tap or paste here to upload images
Cite arxiv.org/abs/2608.00799 in a Space README.md to link it from this page.
Discussion (0)
Sign in to join the discussion. Free account, 30 seconds — email code or GitHub.
Sign in →No comments yet. Sign in and be the first to say something.