New state-of-the-art accuracy of 94.61% achieved on the BANKING77-77 intent detection benchmark through multiview encoder adaptation.

Researchers have achieved a new performance milestone on the BANKING77 intent classification benchmark, reaching 94.61% accuracy on the official test set. This +0.13pp improvement over previous bests was realized using multiview encoder adaptation on the model's final layers, a technique that effectively transferred validation gains to the final evaluation. The model remains highly efficient with a 68 MiB footprint and 216 ms inference time, proving that high-accuracy intent classification doesn't require massive computational overhead or oversized model architectures.