MIT’s New GIFT Framework Teaches AI to Fix Its Own CAD Mistakes

MIT and IBM researchers have built a system that teaches AI models to turn flat 2D drawings into working CAD code, cutting the computing cost by roughly 80 percent compared with existing methods. The framework, called GIFT, also produced designs that matched the geometry of ground-truth models more closely than several competing techniques.

That pitch lands in the middle of a real backlash. Engineers who have downloaded AI-generated 3D parts this year keep running into the same complaint: clean-looking geometry that cannot be edited, simulated or manufactured. GIFT is built for a different category of tool, one that outputs executable program code a CAD kernel can run directly, rather than a static mesh file.

GIFT Learns From Its Own Near-Misses

Vision-language models, or VLMs, take an image and some text and output code. For computer-aided design, or CAD, that code gets executed inside CAD software to build a 3D model engineers can stress-test before anything is manufactured. The MIT team found the real limit on these models is not their size. It is the shortage of diverse, high-quality CAD training data.

Their fix, Geometric Inference Feedback Tuning (GIFT), does not collect new human-labeled examples. Instead it interrogates the model it is trying to improve, using a technique the team describes in four broad moves.

1. Ask the model to generate CAD program code for the same design problem multiple times in parallel.
2. Check each output for correctness by executing it and comparing the resulting geometry against the ground-truth shape.
3. Sort the results into full successes, near-misses and outright failures.
4. Repair the near-misses into working solutions and fold both the repaired and successful outputs into a new training set built specifically around that model’s weak spots.

The paper describing the method, posted to arXiv in March, explains that GIFT combines two mechanisms: soft-rejection sampling, which keeps diverse high-fidelity programs beyond exact matches, and failure-driven augmentation, which turns near-miss predictions into new training examples. Researchers can also cut inference compute by roughly 80 percent once the bootstrapped data is in hand, since the model no longer needs to search as hard at inference time to get a good answer.

If we sample the model 10 times and it generates 10 correct answers to the same problem, then there is not much for it to learn. We care about the in-between cases, where the model might only solve the problem 50 percent of the time.

Giorgio Giannone, the paper’s lead author and a research affiliate in MIT’s Design Computation and Digital Engineering (DeCoDE) Lab who also works as a principal research scientist on Red Hat’s AI Innovation Team, made that point about the in-between cases. Those middling attempts, not the total failures or the easy wins, carry the most useful signal for retraining.

The MIT Team Behind the Feedback Loop

Giannone worked on the project with Anna Claire Doris, a mechanical engineering graduate student at MIT; Amin Heyrani Nobari, an MIT postdoc; and Kai Xu of Red Hat. Akash Srivastava, director of Core AI at IBM and a principal investigator at the MIT-IBM Computing Research Lab, and Faez Ahmed, an associate professor of mechanical engineering who leads the DeCoDE Lab, served as co-senior authors. The work was presented at the International Conference on Machine Learning (ICML) and funded in part by the MIT-IBM Computing Research Lab.

GIFT builds directly on an earlier project from the same lab. In 2025, Doris and colleagues released CAD-Coder, an open-source vision-language model fine-tuned to generate editable CadQuery Python code straight from an image. That project introduced GenCAD-Code, a new training set built specifically for the task. CAD-Coder achieved a 100 percent valid syntax rate and beat general-purpose baselines including GPT-4.5 and Qwen2.5-VL-72B on geometric accuracy. GIFT picks up where that dataset-building effort left off, using the model’s own outputs to keep improving without another expensive round of human annotation.

“Nearly every physical product around us, from airplanes to appliances, begins its life as a CAD model,” Ahmed said of the broader motivation. “Industry teams are eager for AI that can help speed-up the creation of these designs, but today’s models often produce simple shapes inadequate for practice.”

Executable Code Was the Whole Point

That last line from Ahmed doubles as an admission. MIT’s own researchers are naming the exact complaint that has followed AI-generated CAD through 2026: demo parts that look convincing and fall apart the moment someone tries to use them.

Much of that complaint is aimed at a specific class of product. A recent guide from Leo AI, a mechanical-engineering software vendor, argues that most text-to-CAD tools output mesh geometry stored as STL or OBJ files, formats that carry no tolerances, material data or manufacturing metadata. Dessia, another CAD software company, describes the typical AI-generated part as a shape that cannot be edited, simulated or manufactured.

GIFT and CAD-Coder sit outside that criticism by design. Their output is not a mesh at all. It is CadQuery Python code, a script that a CAD kernel executes to build a parametric, editable solid model, the same kind of file format engineers already use for crash tests and durability simulations.

Three Labs, Three Bets on Image-to-CAD

GIFT is one of several research efforts chasing the same problem from different directions this year. Their reported numbers, all self-published by the respective teams, show how differently each group is measuring progress.

Intersection over union, or IoU, scores how closely a generated shape’s volume overlaps the true design. A 12 percent gain on that metric, without any new human-labeled data, is what MIT is pointing to as evidence the self-correction loop works.

Is the Mesh Problem Solved?

Not entirely. Most consumer text-to-CAD tools still output mesh files unsuitable for manufacturing, according to engineering software vendors. A handful of newer platforms generate editable boundary-representation geometry instead. GIFT and CAD-Coder sidestep that argument altogether by producing executable program code.

• Leo AI argues most 2026 text-to-CAD tools remain mesh-based and need heavy rework before they reach a production line.
• Dessia goes further, calling typical AI-generated 3D output a shape with no real engineering substance behind it.
• TheCADHub, a CAD software review site, counters that platforms like Zoo and AdamCAD already generate native B-Rep models in STEP format, not mesh, undercutting the blanket criticism.

The disagreement is really about which tools people are testing. Consumer-facing generators built for speed still lean on mesh. Research systems built for engineering accuracy, including MIT’s, are moving toward code and boundary representations instead.

The Verifier GIFT Still Needs

The GIFT paper is candid about where the method runs out of road. Its authors write that the approach relies on a deterministic geometric kernel and ground-truth CAD programs to check whether a generated sample is correct, which limits its use in unsupervised or real-world settings where no such verifier exists. The bootstrapping phase also carries a one-time offline computing cost before any of the efficiency gains show up downstream.

“With GIFT, we started with geometry because with engineering problems, if the geometry of a 3D shape is not correct, nothing else will be correct, but there are many other aspects to consider,” Giannone said.

That framing matters. GIFT proves a model can be taught to catch its own geometric mistakes when a correct answer already exists to check against. It has not yet been shown to work on a brand-new design problem where nobody knows what the right shape looks like ahead of time.

The Bottleneck Was Data All Along

The CAD software market keeps growing regardless of how this debate settles. One estimate from market research firm Towards Packaging projects the 3D CAD segment will grow to $25.88 billion by 2035, up from roughly $14.31 billion in 2026, as manufacturing, automotive and aerospace firms keep adding AI features to existing design software.

GIFT’s central claim, that the constraint on better AI design tools is training data rather than raw model size, echoes a pattern showing up well beyond engineering software. A separate analysis of enterprise AI rollouts found that corporate bureaucracy caps AI’s returns more than the underlying models do, a similar diagnosis applied to a very different industry.

Giannone and Ahmed say the next step is teaching GIFT to optimize for manufacturability, not just geometric accuracy, and to test the framework on larger models across a wider range of CAD generation tasks.

Frequently Asked Questions

What does GIFT stand for?

GIFT stands for Geometric Inference Feedback Tuning. It is a data augmentation framework, not a standalone AI model, and its bootstrapping phase runs once, offline, before any of the compute savings apply to later use.

How is GIFT different from CAD-Coder?

CAD-Coder is the vision-language model that turns an image into CadQuery Python code. GIFT is the training method layered on top of a model like CAD-Coder afterward, and it works through inference-time scaling rather than retraining the entire network from scratch.

Does GIFT work with any AI model?

The researchers designed it to bootstrap an existing, already-trained VLM without needing a specialized architecture. Giannone has said engineers can point the framework at an underperforming CAD model, set a compute budget, and let the system generate its own training data within that budget.

What is IoU and why does the GIFT paper rely on it?

Intersection over union, or IoU, measures how much a generated 3D shape’s volume overlaps with the true, ground-truth design. GIFT reported a 12 percent mean IoU improvement over a strong supervised baseline, which the team used as its primary accuracy benchmark.

Can engineers use GIFT on real projects today?

Not yet in any packaged form. GIFT is a research framework presented at ICML, and it still depends on a deterministic geometric kernel and ground-truth CAD programs to verify its own outputs, a requirement that does not hold for open-ended, real-world design problems.