A frozen tabular foundation model as candidate generator: propeller-compatible stern designs rise from 1.4% to 99.3%
Synopsis
Engineering design often has only 10 to 1,000 evaluated designs because simulations and experiments are costly. GenPFN trains no new generator: a frozen TabPFN classifier reads labeled designs, candidates are interpolated between desired designs and updated with classifier gradients and Gaussian noise, and every candidate is evaluated by the problem's own constraints and objectives. Across fourteen engineering problems, GenPFN v1 has higher class satisfaction and lower realism distance than the tested CTGAN, TVAE and TabDDPM; in the stern case, 993 of 1,000 candidates passed geometric checks and met the propeller-placement area requirement, versus 28 of 2,000 random context designs.
Interpretation
Augmenting small engineering datasets no longer requires first training a task-specific generative model; candidates can be produced directly under the labels of evaluated designs. Tabular generators such as CTGAN, TVAE and TabDDPM must first be trained on the few available samples, while random sampling can spend much of the evaluation budget outside the desired region. In paired comparisons across fourteen engineering problems, GenPFN v1 exceeds CTGAN, TVAE and TabDDPM in class satisfaction by 69.54, 36.90 and 44.25 percentage points respectively, with significant adjusted comparisons; TVAE gives lower coverage distance and higher diversity, and CTGAN also gives higher diversity.
Generation starts between two designs labeled as the desired class, lets the frozen classifier's gradient push candidates toward higher probability, and uses noise to keep them spread out. Classifier weights stay fixed throughout sampling, and the context supplies only evaluated designs and their class labels, with no refitting for a new problem. Candidates are interpolated in normalized coordinates with one initial perturbation, then receive 100 updates that gate the gradient by a threshold and always add Gaussian noise; in the stern case, sampling 1,000 candidates took 11.97 seconds and geometry evaluation took 0.80 seconds.
After generation, each candidate must be evaluated by the problem's constraints and objectives, and only feasible desired-class designs with their objective values enter regression training. Generating a design vector does not itself supply the objective value regression needs, so every candidate passes through external evaluation before it can augment data. Among observed paired regression fits, adding GenPFN v1 candidates lowers AutoGluon validation mean squared error by 62.1% relative to context-only training (pointwise 95% CI 48.2% to 72.2%), based on 1,960 matching contexts.
Perspective
Evaluation covers fourteen selected engineering problems, including a two-parameter benchmark and a 44-parameter hull problem, plus a hull-stern proof of concept on the row-17 CSOV_X_Mod.csv parent; the stern result describes one parent, one context and one acquisition round rather than an optimization study.
Sampling coefficients are author-selected without systematic hyperparameter tuning; at 100 updates the GenPFN v1 feasibility difference from interpolation plus noise is 1.61 percentage points with a pointwise interval spanning zero, and the adjusted test does not establish a suite-wide advantage; the v3-minus-v1 feasibility difference is 0.0025 percentage points, also spanning zero.