A transformer designed around the Born rule, which connects quantum amplitudes to measurement probabilities, could make one aspect of model explanations mathematically checkable. In an arXiv preprint posted November 3, 2026, Teerawat Tongloy and colleagues report word attributions that sum precisely to a defined prediction score, alongside accuracy comparable to a similarly sized conventional transformer. Peer review has not been verified.
The Born Transformer replaces softmax attention and other standard nonlinear operations with measurement-based calculations. Attention uses the fourth power of the magnitude of query–key overlaps, normalized across each row; a product-based operation replaces the usual feed-forward nonlinearity. The model was trained classically, with two blocks and 1.53 million parameters, approximately 98,000 outside its embeddings.
The explanation guarantee has a specific condition: normalization denominators remain fixed at their values for the original input. Under that condition, the attribution target becomes a polynomial in word amplitudes. Integrated gradients assign each word a contribution, and a 324-node integration rule is mathematically exact for the derived degree bound. Across 1,000 test pairs, the largest reported numerical gap between summed contributions and the class probability above its no-information share was 8 × 10^-12. This guarantees completeness of that decomposition, not a uniquely correct explanation or causal importance in language. Actual word deletion recomputes the denominators.
The researchers tested natural-language inference on e-SNLI: deciding whether a hypothesis follows from, contradicts or is neutral toward a premise. After 20 training epochs across three seeds, test accuracy was 76.98% ± 0.17 percentage points for the Born model and 76.62% ± 0.36 for the softmax baseline. These uncertainties are sample standard deviations, not confidence intervals. A polynomial comparator without the measurement constraints reached 78.98% ± 0.31.
On the shared 1,000-pair attribution subset, deleting the Born explanation’s highest-ranked word changed its prediction in 45% of pairs. Comparable probability-based integrated gradients for the softmax model achieved 32%. The reported difference was 13 percentage points, with a 95% interval of 10–16 points when pairs and seeds were resampled. Any single deletion could change the Born prediction in approximately 53% of pairs. Its evidence overlapped annotator highlights at 0.51, versus 0.46 for comparable softmax evidence. Born attention scored about 0.50 on highlight agreement but only 32% on top-word prediction changes, showing that these measures capture different properties.
Optical compilation reproduced model probabilities within 2 × 10^-15 on 512 pairs in simulation; no optical device was demonstrated. IBM quantum hardware measured attention overlaps for two short pairs, while the remaining computation stayed classical. Predictions and selected deletion responses followed the model, but usable word evidence did not survive the hardware implementation. The work establishes no quantum advantage.
The findings concern one dataset and small models, not large language models. Dataset shortcuts, only three training seeds and dependent comparisons on the same attribution pairs limit generalization. Born probabilities also reacted more strongly to random deletions, complicating probability-drop comparisons. The central result is therefore a conditional mathematical guarantee paired with encouraging benchmark tests, rather than proof that model reasoning has been fully explained.
Cover illustration generated with AI; conceptual, not research evidence.