August 16, 2026
A gradient‑boosted tree beats our 297‑million‑parameter model, including out of distribution
The finding
On RSFQ schema-derivable heads a gradient-boosted tree on tabular features wins by 15–89×, and it wins on the out-of-distribution cell-type split too — which is exactly where the large model was supposed to earn its keep.
We trained a 297.5-million-parameter model to predict the electrical behaviour of superconducting RSFQ logic cells from layout and circuit graph. Then we ran the baseline we require of ourselves: gradient-boosted trees on tabular circuit features.
The trees win by 15–89× on every schema-derivable head.
The part that actually matters
A large model losing to a tabular baseline in-distribution is a familiar story, and the usual defence is that the big model earns its keep when you leave the training distribution.
It does not, here. The trees win on the out-of-distribution cell-type split too — held-out variants, grouped so no clean or mutated version of a held-out cell appears anywhere in training, including self-supervised pretraining, with its own leakage ablation. That is the split the large model existed for.
Ten of the twenty-three served heads now run on gradient-boosted trees in production. It is permanently a hybrid: layout-derived design-rule violations and composite chains have no tabular route, and that is where the learned model is genuinely doing work no baseline reaches.
Why we publish this
Our public model repository ships this baseline in its own results table, above the fold, where it beats us on Ic margin 0.006 against 0.093. A benchmark that only reports the configuration where you win is not a benchmark.
The broader version of this result needs a case pointing the other way, and we have one — on battery curve outputs an MLP beats the tabular baseline by ~13%, which makes the real claim the baseline verdict is task-dependent, which is why running it is mandatory. That half is outside the superconducting stack, so it sits outside what this page covers. Stated here so the scope is visible rather than convenient.
What to take from it
If you are deploying a large surrogate on hardware data, run the tabular baseline on the same grouped split before you believe the architecture is carrying the result. Ours did not, on most heads, and we would not have known without checking.