Lexical Conditioning of Model’s Distribution through Uncertainty-gated Soft-Mixing of Probabilities

Michele Papucci, Giulia Venturi, Felice Dell’Orletta


Abstract
We present Uncertainty-Gated Lexical Decoding (UGLD), a decoding-time framework for fine-grained lexical control in Large Language Models (LLMs) that explicitly addresses the trade-off between controllability and fluency. UGLD adaptively scales intervention through an entropy-based gating mechanism derived from the model’s predictive distribution, activating control when uncertainty is high and limiting interference when predictions are confident. The method supports both promotion toward and against predefined vocabularies. We evaluate UGLD in Italian on two open-weight LLMs (ANITA 8B and Qwen 3 4B) across paraphrasing and free-text generation settings, considering Simple Vocabulary Conditioning and Jargon Reduction scenarios. Automatic evaluation shows consistent improvements in lexical coverage over standard decoding strategies, while human evaluation confirms that fluency is preserved under controlled intervention.
Anthology ID:
2026.readi-1.7
Volume:
Proceedings of the Joint Workshop on Readability and Text Simplification (READIxTSAR) @ LREC 2026
Month:
May
Year:
2026
Address:
Palma, Mallorca (Spain)
Editors:
Matthew Shardlow, Thomas François, Raquel Amaro, Jorge Baptista, Rémi Cardon, Eugénio Ribeiro, Horacio Saggion, Regina Stodden, Amalia Todirascu, Rodrigo Wilkens
Venues:
READI | TSAR | WS
SIG:
Publisher:
ELRA Language Resources Association (ELRA)
Note:
Pages:
89–100
Language:
External URL:
https://lrec.elra.info/lrec2026-ws-readixtsar-07
DOI:
10.63317/5hgiz64xsj6s
Bibkey:
Cite (ACL):
Michele Papucci, Giulia Venturi, and Felice Dell’Orletta. 2026. Lexical Conditioning of Model’s Distribution through Uncertainty-gated Soft-Mixing of Probabilities. In Proceedings of the Joint Workshop on Readability and Text Simplification (READIxTSAR) @ LREC 2026, pages 89–100, Palma, Mallorca (Spain). ELRA Language Resources Association (ELRA).
Cite (Informal):
Lexical Conditioning of Model’s Distribution through Uncertainty-gated Soft-Mixing of Probabilities (Papucci et al., READI-TSAR 2026)
Copy Citation: