Increase generation diversity without sacrificing accuracy

Improve generation diversity in Mahalanobis-Ensemble Decoding while preserving its observed accuracy gains.

Background

Mahalanobis-Ensemble Decoding uses a token-similarity kernel to select a compact, geometry-aware candidate set that balances token confidence and semantic redundancy. The experiments show that it achieves strong accuracy and a favorable accuracy–diversity trade-off, but it does not consistently produce the highest output diversity among all evaluated decoding strategies.

The unresolved problem is to design an enhancement to Mahalanobis-Ensemble Decoding that promotes more diverse generations without eroding the method’s accuracy improvements. The authors identify this as a future research direction, leaving the appropriate mechanism for achieving this balance unspecified.