Paper Accepted at READIxTSAR@LREC2026

I am happy to announce that our latest paper “Lexical Conditioning of Model’s Distribution through Uncertainty-gated Soft-Mixing of Probabilities” has been accepted at the READIxTSAR workshop of Language and Resource Evaluation Conference of 2026 (LREC2026).

The paper is about UGLD, a new decoding method that allow the user to either increase or decrease the relative probability of producing words of a pre-defined lexicons without impacting the LLMs’ fluency. This is done by using an uncertainty-based gating system to the intervention on the decoding probabilities, strenghtening the intervention when the model uncertainty is high.

UGLD will be soon be available on PyPi and freely on GitHub as an open-source project.